Today we’re doing an anonymous Q&A with KL Divergence, a robotics PhD currently in industry working on humanoid robots. For an introduction to humanoid robotics in China, see our article here, and for a deeper look into who’s leading China’s humanoid market, see our latest translated interview with the CEO of Unitree Robotics.
This Q&A covers:
When and how AI-driven robotics will reach a tipping point in viability,
Challenges and solutions for collecting data to build a robotics AI model,
Successful strategies for companies to compete in humanoid robotics.
When will AI + robotics reach a tipping point in viability?
This is extremely difficult to predict.
Here's my non-answer: whenever the world achieves a data flywheel for robotics, i.e. accumulate a dataset large-enough (and algorithms to use it) that allows some robots to achieve a diverse set of somewhat useful tasks, with enough reliability that people allow those robots to operate in their factories, logistics centers, homes, offices, etc.
Once a robot has a “reason for being” in a space, and works well enough, the data flywheel will spin, and the robots will get better and better. This is the same process we are seeing play out in self-driving cars, and why Waymo’s early advantage in deployment is such a big deal, making Tesla dance and move forward plans for Robotaxi. I don't think we will see this happen all at once, in all application domains.
Today, this has already arguably happened for a specific application: robotic picking/packing in e-commerce logistics. Amazon, Dexterity, Covariant, Berkshire Grey, and Ocado all have massive robotics datasets for this basic task, and already use them to create their own “flywheel.” This is short of what we want though, because that data is only for one task, and is specific to those companies' unique robots.
What’s the pathway to viability? How will AI + robotics diffuse through different industries?
So the next stage will likely be doinglots of different tasks (10s-100s) in a structured environment. I would guess logistics centers and manufacturing. I think this could reasonably be achieved in 2-3 years research-wise, and 5-7 years to become commercially commonplace. Along the tail end of that period, you might see these robots start to appear in retail, hospitality, and food service back-of-house. Think: robots doing laundry or restocking shelves. Next, offices. And last, homes.
Will we be seeing AI-driven robots in homes?
We’re 10+ years away even as a question of research, if we ever get there.
Homes (and to some extent offices) are much more difficult than commercial/industrial spaces because of 4 factors: lack of structure and wide variation, safety, and cost.
Structure and variation: Homes are the ultimate “unstructured” environment. They come in infinite variations, and change from moment to moment as people and stuff move around. One day you might decide to put the cucumbers in the top vegetable drawer, then next you might move them into the bottom drawer. Multiply that by everything a home robot might have to ever interact with and the amount of variation becomes mind-boggling. It is impossible to create a system which quantifies and anticipates it all explicitly. The realization has been the impetus for the move towards learned — rather than programmed — robotics AI systems over the last 10 years.
Safety: It’s an engineering achievement to make a robot that can complete tasks and weigh only as much as a smallish human. If that thing falls in a house (dead battery, malfunction, etc.), the stakes are high: it might fall on a pet, break a glass table, or knock over a candle. Contrast with a controlled commercial environment, where people working near the robots can be specifically trained, the environment is arranged so that failures don’t lead to catastrophic danger, and the robots might even be cordoned off behind a cage to minimize the impact of accidents.
Cost: Most proposals for robots in the home have them providing typical domestic labor: cooking, cleaning, tidying up, etc. People already pay for these services in their homes, and it is invariably some of the lowest-paid work in any economy. A humanoid robot has a similar part count and manufacturing complexity to an electric car, so it’s intuitive that the most optimistic cost estimates land the price of these machines at similar numbers: $15-50k, depending on the source. How much would a home robot have to do for a family to justify a $25k price tag with a 5 year life span, assuming no recurring service or subscription costs?
So why do we see some players in the robotics/AI space, humanoid or otherwise, proudly touting their goal of putting robots in homes? My best guess is that it’s a more compelling narrative to attract investment and inspire talent — and it’s not too hard to pivot back to industrial robots anyway.
What are the challenges of getting good training/testing data for AI-driven robots?
You already answered this well in your piece. It's because we need to get robots into the real world to collect lots of good data, but they currently don't work well, can be unsafe, take up space, etc. They have no economic reason to take up space and human attention where you would want them to collect data.
How are proposed solutions addressing these data challenges?
Ways of addressing:
Simulation (but this has flaws, as you mentioned)
Spend a lot of up-front capital to collect robot data directly, in hopes of collecting enough up-front to bootstrap a useful robotics foundation model (i.e. vision-language-action model or VLA)
Find a way to re-use data from the internet, e.g. watching human cooking or furniture assembly videos from youtube (this is very active research, but so far the results have been disappointing),
Master one task at a time (using a combination of 1 and 2, and old fashioned engineering), and hope you collect a dataset diverse-enough before you run out of money (if the dataset is not diverse in tasks, you will have a robot which can do a handful of tasks, but is expensive to train to do new tasks).
What's it like on the ground for a factory collecting this data?
This can vary greatly, but typically you have a task in a factory which has already been defined for a human worker, and it's fairly repetitive. E.g. Tesla's first application is to make Optimus, to grab battery cells rejected into a slide coming out of a battery quality control machine, slot many of them into a grid on a purpose-shaped tote, then walk those totes to a different area of the factory when full. It's very simple and repetitive, and today it's done by humans across dozens of machines. You can imagine other scenarios. For example, sorting packages into bins bound for different geographies in a logistics center. Well-defined tasks and lots of existing automation and built environment (e.g. screens, conveyor belts, well-placed bins and racks) to help humans.
What does it look like to collect the data? That depends on the approach. The most straightforward is teleoperation: a human dons a (usually) VR headset, special gloves for capturing finger movements, and other hardware you'd see in a VR space, and uses them to control the robot directly to do the job. This is “robot-in-the-loop.” It’s slow (the human can't move the robot as fast as they would themselves), and costly (it's actually more expensive than having a human do the job, because it's slower).
Another approach is motion capture: via various methods (camera systems in the work area, body-worn suits, even lightweight worn exoskeletons), we can capture the motion of humans who are already doing the job. This is more speculative, as it’s a difficult research problem to turn these motion recordings into instructions for the robot to achieve the task later.
The last major approach is simulation: usually through the help of a skilled human artist or engineer, create a detailed and functional 3D graphics simulation of the real environment in which the robot is supposed to perform. This allows us to use teleoperation, programmed routines, and reinforcement learning, to control the robot in simulation and collect data on its successes and failures. The weakness of this approach is that the model usually cannot be used immediately on the real robot, because it’s extremely difficult to capture all of the important behavior of a real work task, even a very small one, in a simulation. Roboticists refer to this problem as the simulation-to-reality (sim2real) gap.
On the research horizon, there are a variety of approaches that may allow us to generate or make use of data without actual or simulated robots. A “holy grail” of robot learning for the past decade or more, has been to create a robot learning system which can “learn from watching YouTube videos.” What all of these approaches have in common is that they seek to lower the cost of data for robotics models, by finding ways to make use of lower-quality data (i.e. weaker supervision). The key missing technical piece in most of these approaches is to find a way to map from non-robot behavior in one environment to the actions a robot would take to do the same task in a new environment.
What would indicate a successful humanoid robotics strategy?
How many robots does the competitor have in the real world doing tasks and collecting data, and (importantly) how diverse is that set of tasks? Humanoids are a very expensive way to automate just one thing, so the investment needs to be amortized across many different jobs.
Like bodyguard!
What strategies are robotics firms taking to compete in the market? What will determine who succeeds?
Boy, this is a big question. I won't try to answer the whole thing, but I'll give you a framework.
There are a few fundamental assets to look at here: technical talent (people), chips (compute), robots (how much do they cost? what are their capabilities?), data, and distribution (customer relationships, pilots). Any robotics+AI company or partnership effort needs to assemble all of these ingredients to be successful.
Resources which are less scarce:
Robots: you have a whole article about how China is commodifying robots. However not everyone agrees that *good* robots will be so plentiful (perhaps because of protectionism), and others (e.g. Figure, Boston Dynamics) believe they can create an edge by having the *best* hardware.
Customer relationships: Tech demo deals like the Figure-BMW, Apptronik-Mercedes, and Agility-Amazon partnerships are very low-risk for the larger company and easy to make. CEOs at humanoid companies tell me they have no problem getting 100s of leads.
So a successful strategist will try to gain an edge in the scarcest resources: talent, chips, robots, and data.
Chips: notably — every major NA effort has decided that they need to team up with a giant foundation model provider to have the chips and frontier models to compete. Figure-OpenAI, NVIDIA is in-house, Tesla-xAI, and Boston Dynamics has teamed up with TRI's foundation model team.
Robots: Most people attempting to make their own, however Skild, NVIDIA, and Physical Intelligence have all taken a partnering or purchasing approach for robots instead. Whether a competitor sees robots as a competitive advantage, or an expense, is a major dividing line in strategy in this area.
Talent: immensely cut-throat. Until very recently robotics+AI was a very niche field. An investor told me he believes there are ~25 people in the world who could lead one of these companies well. Even below leadership, the number of people with any training at all in this subfield is in the low 100s. The best-paying outfits in the world with the best reputations take 6 months to hire someone, and are often just waiting for new PhDs to graduate to fill positions. Talent with <1 year of professional experience but relevant education (usually a PhD) can fetch $500k-1M/yr in this field, and/or significant equity, depending on the size of the company.
And finally, data is the most strategic asset these companies seek to accumulate long-term. At the end of the day, the firm who has the best data (or best strategy for getting it) wins the game.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
Hangzhou-based Unitree Robotics is among the top players in China’s robotics industry, developing best-selling civilian quadruped and humanoid robots.
Unitree’s H1 humanoids captivated over a billion people with a traditional folk dance performance during China’s 2025 Spring Festival Gala. Just weeks later, Unitree’s CEO was the youngest front-row participant in Beijing’s highly-anticipated private sector summit.
To better understand the Chinese robotics industry and where it’s headed, we’ve translated and annotated qan interview with Unitree’s founder and CEO, Wang Xingxing 王兴兴. Originally conducted in April 2024 by Titanium Media (TMT), the interview covers:
Why LLMs aren’t enough for the robotics industry, and why Wang predicts the emergence of a large-scale AI model for general-purpose robotics by the end of 2025,
Factors driving the global humanoid robot boom, and why China is uniquely poised to succeed in this industry,
The techno-optimist vision for the economy of the future, powered by humanoid robots as well as machines of alternative forms,
The timeline for mass adoption of AI-powered general-purpose robots,
Unitree’s strategy for competing against foreign and domestic robotics firms.
We’ve added some editorial notes for your enjoyment, including commentary by anonymous robotics PhD and current industry player KL Divergence.
Original Article | Archive | Title: “Dialogue with Wang Xingxing: Humanoid Robots Will Reshape All Industries Within My Lifetime” | Author: Rao Xiangyu 饶翔宇 for Titanium Media (TMT) 钛媒体APP | Editor: Zhong Yi 钟毅
On February 17, 2025, a highly anticipated private enterprise symposium was held in Beijing.
At the event, the presence of Wang Xingxing, the founder of Unitree Robotics and a post-90s entrepreneur, attracted market attention. Wang was seated in the front row among the business representatives, alongside industry giants such as Zeng Yuqun, Jack Ma, Ren Zhengfei, Wang Chuanfu, Lei Jun, and Pony Ma. Among them, Ren Zhengfei, Wang Chuanfu, Liu Yonghao, Yu Renrong, Wang Xingxing, and Lei Jun delivered keynote speeches.
As a startup company, Unitree Robotics and Wang Xingxing have experienced what can be described as a “rocket-like leap” in growth.
[KL Divergence:Perhaps. Though consider that Unitree has been an established player, particularly in quadruped robots, for 10 years. They have worked hard and scraped their way up inch by inch by serving the small markets which existed for their product (mostly researchers). Their recent prominence has come from riding the wave of interest in humanoid robots by creating low-cost, easy-to-use, but not particularly advanced or capable, humanoid robots for research.]
Public records show that Wang Xingxing earned his bachelor's degree from Zhejiang Sci-Tech University. Due to poor English proficiency, he failed to gain admission to Zhejiang University for his master's studies and was instead placed at Shanghai University. In an interview, Wang once mentioned, “During my three years in high school, I only passed English exams three times in total.”
From 2013 to 2015, while pursuing his graduate studies, Wang, despite having limited resources and funding, independently designed hardware and control algorithms and combined them with industrial motors to develop the robotic dog XDog. This project won second prize in the Shanghai Robotics Design Competition. After graduating, Wang embarked on an entrepreneurial journey focused on robotic dogs.
[KL Divergence: Actually, much of the hardware and control algorithms were from publicly-available robot and actuator designs published by Western researchers, such as Sangbae Kim (MIT) and Dan Koditschek (UPenn). What Unitree really excelled at was (1) iterating high-performance actuators, robot designs, and research-grade control algorithms, and (2) leveraging the Chinese supply chain to create low-cost, high-performance, highly-reliable combinations of these key technologies. In other words, they productized the research.]
Unitree Robotics, founded in 2016, initially specialized in the development of quadruped robotic dogs and successfully sold them worldwide, becoming one of the leading players in the industry in terms of product shipments. By 2023, the company ventured into humanoid robotics and quickly became one of the most closely watched companies in the field. In 2025, Unitree Robotics’ latest humanoid robot appeared on the stage of the CCTV Spring Festival Gala, garnering widespread public attention.
In April 2024, Wang Xingxing, the founder of Unitree Robotics, had an exclusive interview with Titanium Media APP. The relevant content can be found below.
Wang told TMT APP that the fundamental reason behind the humanoid robotics boom is the emergence of large AI models. Previously, it would take one to two years for a humanoid robot to learn to walk, but now, with AI algorithms, this can be achieved in just one month.
Regarding the future development of humanoid robots, Wang expressed strong optimism. He believes that by the end of 2025, at least one company worldwide will have developed a general-purpose robotic AI model. This foundational model, he explained, is like a complete set of building blocks, with large language models being just one piece. Other crucial components include visual perception, tactile sensing, decision-making, and interaction systems.
Looking at an even longer timeline, Wang told TMT, “Within our lifetime, humanoid robots will be able to revolutionize every industry, from manufacturing and agriculture to services and industrial sectors. Taking it a step further, governments could potentially deploy 100,000 humanoid robots to construct an entire city. On a smaller scale, robots could even shrink down to the size of cells, transforming all aspects of our natural environment.”
Below is the full interview between TMT and Unitree Robotics founder Wang Xingxing, with slight editorial adjustments.
“The Turning Point for Humanoid Robots Has Not Yet Arrived”
TMT: A few days before our meeting, Boston Dynamics, a star company in the robotics field, announced that its hydraulic-powered humanoid robot would be phased out, and future developments would focus on electric-powered products. What are your thoughts on this?
Wang Xingxing: Boston Dynamics has been making robots for many years, and they’ve also been working on commercialization for a long time.
As for hydraulic drive systems, I had already believed before 2013 that this approach could not be commercialized. The reason is simple: it relies entirely on precision mechanical components, and once you involve such components, costs will never come down. Moreover, all hydraulic systems leak oil. That’s why you hardly see hydraulic systems in consumer vehicles anymore — they’ve all been replaced by electric drive systems.
So, if Boston Dynamics wants to continue developing humanoid robots, switching to electric drive is definitely the right path. The only surprising thing to me is that I assumed around 2018 that they had already started working on an electric version. But later, when they had made no detectable progress, I just stopped paying attention.
TMT: Compared to hydraulic systems, is electric drive better suited for large AI models?
Wang Xingxing: Compared to hydraulic systems, electric drive is all advantages and no disadvantages. As for whether electric systems are better suited for AI, that’s harder to say. However, electric drives have lower production costs, offer greater motion flexibility, are safer, and also lighter in weight.
TMT: Now that Boston Dynamics has switched to electric-powered robots, combined with their existing training data, do you think they could iterate faster than competitors in this new wave of competition?
Wang Xingxing: It’s hard for me to say. However, we remain quite confident, because we’ve been working on quadruped robots for many years, and a lot of the algorithms and components we’ve developed can be directly applied to humanoid robots.
Another important point to note is that most of the top AI talent in the U.S. isn't at Boston Dynamics—they’re at Google, NVIDIA, and OpenAI. Boston Dynamics' strength likely lies in hardware development and traditional humanoid robot control systems.
TMT: So, would you say that the emergence of large AI models is a major turning point for humanoid robots?
Wang Xingxing: I don’t think we’ve reached that turning point yet.
Right now, it’s more like a starting direction. There's a common misconception—many people think that large language models like ChatGPT can be directly applied to robots, but in reality, that’s not the case.
TMT: Why not?
Wang Xingxing: Because LLMs aren’t designed for robotics in the first place. ChatGPT operates purely on text logic, and its entire training dataset is based on text data. It doesn’t perform well in robotic environmental perception — this is a global challenge, not just a problem for one company.
While the humanoid robotics industry does use AI, the technology is actually very different from large model technology.
TMT: But some companies have claimed that large AI models can already recognize different types of plates, allowing robots to identify and pick them up.
Wang Xingxing: That’s not something we can verify. It was just a video, and no one has confirmed its authenticity.
Besides, there’s no data proving that if you swap the plate for an apple, a pear, or something else, the robot would still be able to recognize and handle it correctly. Personally, I don’t see any evidence of real technical breakthroughs coming from Silicon Valley — it still seems quite conventional (中规中矩).
TMT: So, large AI models are not the key turning point for humanoid robot development? Are they less important than people think?
Wang Xingxing: The models themselves are not important for robots, but the underlying technological direction they represent is very important.
Right now, large models are mainly focused on language models — but no one has yet developed a true large-scale model specifically for robotics.
TMT: That brings us to the big question — what triggered the humanoid robotics startup boom in 2023?
Wang Xingxing: The reason is really quite simple: Tesla started working on humanoid robots.
Elon Musk has already disrupted industries like automobiles and rockets, growing them into massive sectors. Now that he's entering humanoid robotics, governments and various institutions want to get started early, rather than waiting for Musk to succeed first and then trying to catch up.
[KL Divergence: I think this is a little bit just-so and playing to the audience a bit too much. The fundamentals are more important. Elon and Optimus is definitely the spark which ignited the wildfire. But the kindling was years and years of slow and steady progress on batteries and electric motor power density made it finally possible to create practical (as in, strong and light enough) electric humanoid robots, around 2020-21. Elon's team caught on to this a little early, because these are also technologies that Tesla happens to to be deeply invested in. But others were doing it already, just quietly.]
At the same time, ChatGPT and other LLMs have expanded the public’s imagination of AI’s potential. You could say these models ignited excitement and enthusiasm across the industry. Right now, what we’re seeing is just the beginning — the momentum will only grow stronger.
As hardware and AI technology advance each year, the impact of humanoid robotics on the world will be massive and transformative.
“It’s simple, really not as complicated as most people think”
TMT: Current large AI models are just the beginning. What's the future direction of the industry or where should everyone's efforts be focused?
Wang Xingxing: There are many directions. The first step is adapting AI for robots - developing robotic vision, perception, understanding, execution planning, and various operations.
I'm excited just like everyone else. I personally feel this industry will develop rapidly, including robots, large models, and AI. I believe by the end of 2025, at least one company globally will develop a relatively general-purpose robot large model.
Our company hopes to develop it ourselves, but realistically speaking, the probability is higher that an American company will achieve it first.
[Angela: Wang is optimistic. Depending on what goalposts you set, training a robot “foundation model” requires large, multimodal datasets that take time and capital to collect or synthesize – including vision, sound, touch, motion, social and environmental interaction, and so on. In some ways, Wang’s prediction aligns well with the Chinese government’s stated goal of mass production of humanoids by 2025 and world leadership by 2027. At the same time, he realistically recognizes the strength of US innovation. The main takeaway here is that to Wang — and likely to many of his industry counterparts — the humanoid robot race is accelerating towards some decisive moments.]
TMT: So that brings up the question of open source versus closed source.
Wang Xingxing: If we develop it, it definitely won't be open source.
TMT: Is there a unified model between robot large models and robot dogs?
Wang Xingxing: Most robot dogs are implemented through reinforcement learning, which is a relatively mature technology.
Robotic large models or robot world models can be applied to all robots, not necessarily humanoid or dog-shaped ones - they're universal tools. I've always believed that robots don't necessarily have to be humanoid; the humanoid shape is just one of many possible forms. I've never insisted they must be humanoid.
TMT: But the mainstream view is that humanoid forms are better because our whole society is built for human-shaped frames.
Wang Xingxing: They might like to say that, but I've never believed it.
You can build entirely new physical worlds. Why would you need a humanoid form for mining? Why would you necessarily need a human shape for building houses? Of course, humanoid forms are important, or relatively important, but they're not everything.
For example, at home, people might prefer humanoid robots for performing scenarios or accompanying you on trips. But for building houses or transporting things - physical labor - there's no need for them to be humanoid. Plus, humanoid forms might give people a sense of owning a slave if you make them do unpleasant work, making their owners feel uncomfortable.
TMT: Would you feel sorry for them?
Wang Xingxing: Current AI hasn't reached that level yet; it can’t perceive such things.
But if its AI could perceive pain or negative emotions, then yes, that might be problematic. But there’s no need to feel sorry now, because it’s still just an inanimate object [死物, literally a ‘dead thing’] with limited intelligence.
TMT: I'm curious about something — even though their intelligence is limited, when you push them, why do they display human-like staggering movements?
Wang Xingxing: Because that’s what the AI was trained to do through reinforcement learning.
TMT: So it’s imitating human behavior?
Wang Xingxing: Some behaviors aren't imitation; they’re determined by natural laws. You could say physical laws constrain these robots to move in certain ways. If an alien had a human shape, its movement would probably be the same as well.
TMT: Currently people break down robots into cerebrum, cerebellum, and the physical body. What's your view on this?
Wang Xingxing: I’ve never liked separating the cerebrum and cerebellum so distinctly. One model is enough - why divide it into two? I don't think it's necessary.
Of course, there might be various modules within the model, but overall I prefer treating it as a single model. From walking to fine-grained operations, we implement everything using AI in a completely end-to-end manner. From visual perception to leg execution, one model handles it all - no intermediate mathematical formulas whatsoever.
TMT: Can the hardware capabilities keep up?
Wang Xingxing: For robots, it's just a few joints — it’s really not that hard. Just sensors feeding into the model, and then the model outputs to the joints. That's all.
TMT: Your understanding of humanoid robots seems simpler than others'.
Wang Xingxing: It is simple, not that complicated.
TMT: For example, others might think dexterous hands are difficult for fine operations because they require more accurate recognition and finer motion control.
Wang Xingxing: It's very difficult if you use traditional technology, so you can't rely on traditional approaches. Without technological innovation, there's no point in working in this field. Of course, you can't express it so directly — better not to go too far beyond public understanding, otherwise I'd probably get cursed out (骂死).
TMT: What specifically do you mean by non-traditional?
Wang Xingxing: It's new AI, end-to-end. It means not having to manually write lots of software programming rules in between, nor perform traditional image recognition.
TMT: How do you implement that?
Wang Xingxing: Modify the model. The underlying AI is the same, but your entire model structure and algorithms are different. I can't explain this too specifically - it would be hard to understand. For example, you don't need traditional image annotation or image understanding at all. You can input images and videos into a model, and the output is directly the robot's joint trajectories, then you just train it. You can still do image annotation, like labeling images of apples. But annotation has only one function: interacting with humans, helping it better understand people. For the robots themselves, there's no difference between an apple and a pear.
TMT: Compared with the mainstream opinions, your logic and industry judgments are unique.
Wang Xingxing: The mainstream viewpoint still has many issues. As a startup, if your thinking is just mainstream, it just won’t work out well for you. You must see the development direction for the next few years, and once you see it, plan ahead accordingly - then you're certain to win, or at least not lose. If you only see what everyone else is talking about, others can certainly do better than you - how could you stand out?
TMT: In your view, what will the next few years look like?
Wang Xingxing: I can't get too specific, but what's certain is that the industry will progress extremely rapidly.
TMT: How fast are we talking here?
Wang Xingxing: It's basically beyond imagination. The current pace of AI deployment in factories — globally, technological progress is extremely fast and has almost proven viable.
TMT: Currently, no company can fully utilize robots for work.
Wang Xingxing: But the entire logic has almost been proven. This doesn't mean robots can do everything, but work-capable, end-to-end robots are nearing maturity. A more general-purpose robot model will likely be developed by a company globally before the end of 2025.
TMT: That fast?
Wang Xingxing: It could be even faster. Some people have already seen where this is going - though it sounds a bit boastful, I feel I've seen it too. Following this direction, with some additional time, manpower, and money, it can basically be developed.
“All technological breakthroughs have a large element of luck"
TMT: What specifically does a robot model refer to?
Wang Xingxing: You can think of it this way: first, it has strong mobility capabilities applicable to most terrains, possibly with some mobility skills exceeding humans. For instance, its obstacle-crossing ability, speed, jumping ability might be better than humans. Another aspect is working in factories - it can do many tasks without requiring manual programming. Through large model capabilities, with just a little teaching, it can learn by itself and then perform well.
TMT: Is simulation training in virtual environments still necessary?
Wang Xingxing: Probably not all that necessary. Once you've trained it well and validated it, you don't really need simulation anymore. Of course, completing the hardware won’t happen right away, but I think that's just a matter of time. As for AI, there's still some uncertainty. Although I just said I'm personally optimistic it will emerge before the end of 2025, it might not happen - it could take 3-5 years before it's developed. It depends on humanity's collective luck - sometimes it just comes down to luck.
TMT: How do you understand this kind of luck?
Wang Xingxing: Many technological breakthroughs depend on luck. For example, if Einstein hadn't existed, someone else would probably have discovered his theories. But it might have been delayed by several years, or even decades. You can consider that all technological breakthroughs have a large element of luck involved.
TMT: Another point: besides algorithms and models, large models need data. Is data collection currently very difficult?
Wang Xingxing: There are indeed many things that need to be done, but there are methods for addressing them. It's not as complicated as people think - many problems aren't as complex as people imagine. You know, in all current technology fields, if you really look, there's nothing truly complex; everything is relatively direct and simple. Even
TMT: So is your industry also divided into two camps - optimistic and pessimistic? For example, you're more optimistic, thinking the whole thing isn't that difficult.
Wang Xingxing: It definitely requires time and intellectual investment, but these are things that can be solved and advanced. They’re not like room-temperature superconductors or controlled nuclear fusion. The biggest problem with room-temperature superconductors and controlled nuclear fusion is that there’s a question mark over whether they’re physically possible. The universe might simply not allow such things to exist, and humans might never achieve them no matter how much time and effort we invest.Artificial intelligence robots are common things, not something extraordinary — just the intelligence of a bunch of humans and animals. Intelligence is a widespread phenomenon. Some animals are very smart and can understand much of what humans say, they just can’t speak. And crows — some crows can even use tools directly.So, intelligence doesn’t have many limitations or physical constraints; it can be replicated.
TMT: What’s the biggest motivator for your work?
Wang Xingxing: To be honest, what moves me personally is AI.
A few years ago, an investor asked me whether our company would ever develop humanoid robots, and I told him, “We would never do it, even if it kills us.”
[KL Divergence: Great honesty here. It's true. Virtually the entire field considered humanoid robots a hopeless tarpit, which would consume all of your time and money and render not progress. Even in robotics research, humanoids were a quirky backwater reserved for the cranks and over-optimistic.]
Back then, humanoid robots were far too complex. Traditional algorithms simply couldn’t handle such intricate machines. The conventional approach to training humanoid robots relied on highly skilled engineers manually writing mathematical equations to model movement. These equations would then be solved to determine the robot’s motion trajectory. But this method had severe limitations—if the environment changed, the equations often became invalid, requiring entirely new models to be designed from scratch.
This approach also led to an overwhelming amount of code, and as the system grew more complex, it became nearly impossible to maintain manually. However, AI changes everything. As long as the model is well-structured and continuously fed with data and compute, AI can iteratively optimize itself through trial and error. By leveraging reinforcement learning and reward mechanisms, AI can automatically retain successful training outcomes and discard ineffective ones, dramatically improving training efficiency.
Recent progress in AI technology—both in capability and speed—has far exceeded my personal expectations. That’s why, despite having worked on humanoid robots for just over a year, our performance is already exceptionally good. The reason we’ve been able to move so quickly is simple: thanks to advancements in AI.
The benefit of AI is that once you’ve built a strong model, the rest is just a matter of compute—you don’t have to manually fine-tune everything. If you need to test a scenario, OK, all you need to do is feed the system more data. This is also why Tesla’s autonomous driving team is significantly smaller than Chinese autonomous driving teams. I know for a fact that Tesla’s team has only a few hundred people, whereas some companies in China have teams numbering in the thousands.
TMT: Is this also why newer players have been able to surpass Boston Dynamics?
Wang Xingxing: Exactly. If we were competing with Boston Dynamics purely using traditional algorithms, we wouldn’t stand a chance. The reason is simple: Boston Dynamics has an entire team of PhDs from MIT, and there’s no way China could outmatch them in that domain.
TMT: Looking ahead, what do you think will be the key differentiator among humanoid robots?
Wang Xingxing: Robotics is an integrated product. Unlike fuel-powered vs. electric vehicles, where the underlying technologies are fundamentally different, the differentiators in humanoid robots will be more incremental—primarily in specific engineering optimizations, such as motor scale, motor placement, workspace dimensions, structural design, and leg configurations.
The same principle applies to AI. Take large language models — they’re fundamentally pretty similar. The biggest points of differentiation are in the details rather than in fundamental design; OpenAI’s GPT architecture is still relatively clean.
“In our lifetime, humanoid robots can reinvent all industries and the natural environment.”
TMT: Commercialization is also important. How can startups survive in an increasingly competitive landscape?
Wang Xingxing: The business logic is very simple. As long as your product is better than your competitors’ in all dimensions, then you will profit. What remains is the question, how big is the entire market? Right now, our company has a strong market position, so we have captured most of the easily accessible revenue opportunities.
TMT: What do you mean by ‘easily earned revenue’?
Wang Xingxing: From having high shipment volumes. We sold quite a few quadruped and humanoid robots last year.
TMT: How many did you sell?
Wang Xingxing: It's hard to say exactly, but it's under a few hundred. However, we definitely sold the most in the domestic market.
TMT: Who bought them?
Wang Xingxing: A variety of buyers, including research institutions, AI companies, and businesses pursuing real-world applications.
TMT: How can you move so fast and manage to sell your products?
Wang Xingxing: Because we have a strong foundation. There’s significant overlap between robotic dogs and humanoid robots. Our company holds advantages in technical R&D, AI algorithms, manufacturing, and sales channels. We already have an established customer base and ready-to-market products. Other companies have to build everything from scratch, which takes time.
TMT: Is your revenue sufficient to support R&D?
Wang Xingxing: Our company maintains healthy gross profit margins, complemented by ongoing funding.
TMT: For humanoid robot startups, is the ability to secure funding a core advantage?
Wang Xingxing: It’s hard to judge the industry right now because it’s too hot. Many companies with basic foundations have raised some funds, which are at least enough to keep them afloat.
There’s certainly no shortage of funds in this industry. When we started, we were poor. Compared to back then, things are completely different. Now, some companies have been around for only a year and already have a valuation of 1 billion yuan. It's astonishing. The industry isn’t short on capital, and neither are they.
But I think that before the industry truly takes off, having too much money is pointless. It’s difficult to allocate effectively, and if spent indiscriminately, it could easily be wasted. At this stage, neither the technology nor the business models have been fully validated, so throwing money around wouldn’t be wise.
Take bike-sharing, for example. It worked because the business model made sense. Once that’s proven, the only thing left is scaling up, and there’s nothing left to do but pour in funding.
TMT: What do you mean when you say the technology and business model haven’t been fully validated?
Wang Xingxing: It means that neither the technical framework nor the commercialization strategy is fully developed. Even if you have the capital, you don’t necessarily know how to deploy it effectively.
TMT: What are the main technical challenges?
Wang Xingxing: For humanoid robots, the biggest question is how to integrate with AI models—we don’t have a definitive answer yet.
TMT: Another observation—most humanoid robotics entrepreneurs today are quite young. (Wang Xingxing is from the ‘90s generation.) Why is that?
Wang Xingxing: It’s simple. Older generations just aren’t as interested in this space. AI technology is evolving at an unprecedented pace and older knowledge is becoming outdated – knowledge of the technology we had five years ago is practically irrelevant. The younger generation is fastest at learning and applying the new advancements. Traditional internet startups had a low barrier to entry — basically anyone could become a product manager. But humanoid robotics isn’t a conventional industry.
[Angela: We’ve written before about how this generation of emerging technology creates space for young, enthusiastic talent to make an impact — DeepSeek is a good example of this. It would be interesting to know if, like DeepSeek, Unitree draws its success from China’s pool of homegrown talent. Wang Xingxing himself never studied or worked abroad.]
TMT: Earlier, you mentioned the potential for a breakthrough innovation. Were you referring to how humanoid robots and AI models can be integrated?
Wang Xingxing: Yes, more or less.
TMT: But aren’t AI models just modular components that can be put together like building blocks?
Wang Xingxing: The differences in AI models go far deeper than that. Take Transformer architectures, for example—there are still endless ways to optimize and refine them. Researchers are even exploring alternatives to Transformer-based models altogether. The AI field is full of opportunities for technical breakthroughs, and there’s still vast room for innovation.
I anticipate that by 2025, we’ll see a significantly improved AI model for general-purpose humanoid robots. When that happens, industry momentum will accelerate even further, to the point where companies from around the world try to enter.
TMT: At that point, do you think hardware or software will be the first to breakthrough?
Wang Xingxing: Software will be the key driver. No matter how advanced the hardware is, without the right software, it’s just an expensive pile of metal.
TMT: So given the current pace of development, as soon as the right software emerges, the hardware will be able to keep up?
Wang Xingxing: Absolutely. Hardware won’t be a bottleneck. If it’s really needed now, we can scale production quickly by aggressively deploying capital. If necessary, we could push manufacturing capacity to its limits — pay engineers 10 times their normal salary, work around the clock, and purchase all the necessary equipment. With sufficient investment, we could have mass production up and running in as little as a few months to a year.
TMT: How does China’s hardware capabilities compare to those of other countries?
Wang Xingxing: China has a significant edge in hardware. The cost-performance ratio is much higher.
TMT: Why is that?
Wang Xingxing: First, in the U.S., hardware development doesn’t receive as much attention—most of the top talent is focused on software. Second, manufacturing and labor costs in the U.S. are much higher than in China.
TMT: It seems like they are prioritizing software, while our strength lies in hardware.
Wang Xingxing: Exactly. Most major U.S. companies focus primarily on software. But at Unitree Robotics, we are developing both software and hardware, because maintaining competitiveness requires full-stack capabilities. As a relatively smaller company, we can’t afford to focus on just one domain. Large corporations can get away with specializing in software and outsourcing hardware, but for us, abandoning hardware development would be an unwise strategic move.
TMT: Why has robotic dog technology matured faster than humanoid robotics?
Wang Xingxing: One reason is that robotic dogs have been in development for a longer period, and their form is more stable. They don’t require complex dexterous manipulation, like grasping and handling objects.
Another key reason is that, in the past, there was a much larger community of developers working on robotic dogs, whereas today that number has declined. In AI, the maturity of a technology is often directly correlated with how many researchers are actively working on it.
For example, large language models have advanced much faster than AI for robotics simply because more people are involved in developing them. Ten years ago, computer vision—especially facial recognition—was in its golden age because so many researchers were working on it. Image-based AI took off because it was relatively easy to experiment with; all you needed was a decent computer.
But robotics is a different story. It requires hardware simulation and real-world testing, which makes it much harder for individuals to participate. That’s why the field has been slower to progress. However, as I mentioned earlier, the industry is now accelerating because a growing number of people are entering the space. More minds working on a problem naturally lead to faster breakthroughs.
TMT: Does Unitree Robotics have a clear product roadmap and timeline?
Wang Xingxing: We will launch new products every year.
TMT: What do you envision for Unitree’s next-generation robots?
Wang Xingxing: The next generation will undoubtedly surpass current models in every aspect—appearance, performance, AI capabilities, and more.
TMT: Can you give a specific example?
Wang Xingxing: Our goal is for humanoid robots to perform real industrial tasks—working in factories, assisting in production assembly, and handling logistics.
TMT: Do you have a release timeline for the next generation of robots?
Wang Xingxing: It’s not convenient to disclose at the moment.
TMT: Unitree Robotics has already completed eight rounds of funding. Do you expect fundraising to accelerate moving forward?
Wang Xingxing: I think we’ll be fine. As the industry gains more attention, we’re seeing increased interest from investors.
TMT: What do you think will be the ultimate future for humanoid robotics?
Wang Xingxing: In the future, humanoid robots could redefine entire industries—from manufacturing and services to agriculture, mining, and construction.
I imagine a distant future in which governments could deploy tens of thousands of humanoid robots to build entire cities from the ground up. At that point, infrastructure is fully automated, housing is provided at no cost, and people no longer need to work because robots sustain the entire economy. That’s entirely within the realm of possibility.
Also, right now when we talk about humanoid robots, we picture machines that are roughly human-sized. But in reality, humanoid robots could build smaller robots, and those smaller robots could build even smaller ones. This process could continue indefinitely, leading to robots at microscopic scales.
Eventually, we might see robots as small as biological cells. Who knows what’ll happen then? What we perceive as bacteria could actually be tiny robotic entities. The entire natural environment could be restructured from the ground up. When that happens, governments will need regulations to prevent unchecked proliferation, or these robots could consume resources uncontrollably.
[Angela: A very science-fiction vision indeed. But Wang’s fantasy of a robot economy resonates with Beijing’s investment in industrial robotics as a path for economic advancement. ChinaTalk will continue tracking such developments in robotics.]
TMT: Do you think we will see this level of technological advancement in our lifetime?
Wang Xingxing: Absolutely. The only missing piece is AI. Once AI breakthroughs happen, everything else will follow naturally.
This will fundamentally reshape the world. I’ve always believed that when we look back at today’s society after the emergence of general-purpose AI and humanoid robots, it will feel as distant and primitive as looking back at the Stone Age.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
Announcing: the ChinaTalk book club! We have upcoming shows with the authors of To the Success of Our Hopeless Cause: The Many Lives of the Soviet Dissident Movement, To Run the World: The Kremlin’s Cold War Bid for Global Power, andLearning by Doing: The Real Connection between Innovation, Wages, and Wealth. We’d like to encourage you to read along with us in preparation for the shows!
Manus, a Wuhan-developed AI agent went viral this weekend. Guests Rohit Krishnan of Strange Loop Canon, Shawn Wang of Latent Space, and Dean Ball of Mercatus and Hyperdimensional join to discuss.
We get into:
What Manus is and isn’t,
How China’s product-focused approach to AI compares with innovation strategies in the West,
How regional regulatory environments shape innovation globally,
Why big AI labs struggle to build compelling consumer products,
Challenges for mass adoption of AI agents, including political economy, liability concerns, and consumer trust issues.
Jordan Schneider: This past Friday, Monica — a startup founded in 2022 — launched a product called Manus. The launch was done through a video in English. Manus is ostensibly an AI agent that you tell to do something, and it can interact with the internet as if it were a person clicking around to book a restaurant, change a reservation, or potentially one day take over the world through Chrome. The rollout was remarkable, with hype building dramatically over the weekend. The product seems to be more competitive than similar offerings we’ve seen from OpenAI and Anthropic. With that context, Shawn, what were your first impressions of what Manus was able to build?
Shawn Wang: My first impression was that it’s a very well-executed OpenAI Operator competitor. It can effectively browse pages and execute commands for you. In side-by-side tests that people in the Latent Space community were running with Operator versus Manus, Manus consistently came out on top. This is backed by the benchmarks they hit on the Facebook Gaia benchmark, which evaluates agents in the real world [which is a public benchmark]. The product is very promising. I’m somewhat suspicious about how well the launch was executed with influencer-only invite codes and people writing breathless threads. We’ve seen this many times in the agent world, but this one people actually seem able to use, which is nice.
Rohit Krishnan: What interested me most was that all our previous conversations about China focused on models — how good their models are, how much money they have, how many GPUs they possess, etc. Now we’re talking about a product. The closest thing to a product from DeepSeek was their API, which is really good with an exceptional model, but the interface was just the same old chat interface. We’ve been discussing agents extensively for a long period. In the West, we still live under the umbrella of fear regarding AI agents, which is why most models aren’t given proper internet search capabilities. It’s amazing to see that the first really strong competitor has come out of China — arguably better, perhaps slightly worse, but definitely comparable to Operator. They made it work with a combination of Western and Chinese models, using Claude and fine-tunes of Qwen underneath. That changes the product landscape as far as I can see.
Dean Ball: I didn’t get an invite code myself but was able to use someone else’s account briefly. I ran it through my favorite computer use benchmark that I’ve organically discovered — trying to book a train ticket on amtrak.com. Operator consistently fails at this task, but Manus succeeded on its first attempt. That says something significant! Many other demos I’ve seen of the product seemed quite impressive and like things that would surprise me if I saw Operator doing them, based on my perception of Operator’s reliability and competence.
Humanity's Last Exam
This isn’t a story about some shocking technological innovation or about DeepSeek’s unfathomable geniuses, as Jack Clark says, discovering some new truth about deep learning. This is good product execution in a style relatable to Y Combinator circa 2015. It’s a well-built product that works effectively, though it has flaws and glitches like all computer use agents do. What’s interesting is that this represents an advancement of capabilities. DeepSeek might be a more impressive technical achievement than R1 and V3 in some fundamental sense, but DeepSeek R1 wasn’t better than OpenAI o1 — at best, it was comparable. Manus appears better than what I’ve seen from Operator, or at minimum comparable but I think unambiguously better. Thinking about why that’s the case is a really interesting question.
Jordan Schneider: Shawn, you just hosted a wonderful week-long AI agent conference in New York City. What’s your take on why no one in the West got to this first?
Shawn Wang: There are many skeptics of agents, even among the agent builders themselves. People range across the spectrum in terms of the levels of autonomy they’re trying to create. The general consensus is that lower levels of autonomy are more successful. Cursor being an agent is now worth $10 billion, whereas the people who worked on Baby AGI and Auto-GPT are no longer working on those projects.
Working on level four or five autonomy agents hasn’t been a good idea, while level one and two — more “lane assist” type autonomy agents — has been the better play. With the rise of reasoning models and improvements in Claude and other systems, that is changing every month. The first one to get there, like Manus, would reap the appropriate rewards.
Dean Ball: That intuition you’re describing is definitely something I’ve heard too, and it’s probably right. For a variety of reasons, I won’t be using Manus on a day-to-day basis. Part of that involves security concerns, but even without those concerns, I’m not sure this is a product I would use regularly compared to an agent like OpenAI Deep Research or a Cursor-style product. Those have much more genuine day-to-day utility.
As an investor in this company, I would be concerned that Manus will be, to use Sam Altman’s terminology, steamrolled by the next generation of computer use agents from the big labs. That’s very possible. From a practical business and technological perspective, this makes sense to me.
Rohit Krishnan: The key question I keep pondering is why Manus wasn’t built by a YC company six months ago. We’ve internalized the fear Dean talked about — that anything we build will get steamrolled by Sam Altman. In some ways, that’s correct. We all personally know code assist companies that emerged a couple of years ago and went bankrupt when the big labs effectively took over.
However, I have this heretical notion that despite everyone talking about agents, nobody at the large labs cares enough about them. They don’t seem interested in building products beyond making models smarter and letting them figure out products on their own. We’re stuck in this weird situation where I have access to every large model in the world, but half of them can’t do half the things because nobody has prioritized those capabilities.
O1 Pro can’t take in documents. O3 Mini couldn’t take in Python files or CSVs. Claude can’t search the web. These weird restrictions exist partly from AI safety concerns and partly because nobody has bothered to add these features.
One significant benefit of something like Manus is that people are actually trying to build useful agents for real-life tasks, like booking Amtrak tickets — which is a great evaluation benchmark. This pushes the labs or anyone else to say, “We should probably try to do this.” We can’t just throw up our hands and wait a year hoping the labs will build the next big thing.
The Western success story is effectively Perplexity — the one company that did what the labs would have been closest to doing but never did, and found success. Beyond that, when thinking about other agents we normally use, I can only realistically name Code Interpreter from a couple of years ago and Claude Code, which just released. Both are stripped-down versions that do a few things but still can’t handle basics like search.
When I look at Manus, what stands out isn’t just that they made an agent ecosystem work using external or combined models, which everyone expected would happen. More importantly, they actually went for it. There’s a price to pay — you have to try it. You can give it browser access, let it work for four hours, and get something useful back. Unlocking this capability is important from a product perspective.
Jordan Schneider: I’ll give one more perspective as well, which could be a fun US-centric observation. In the US, we’re very interested in B2B and developer tooling, especially in Silicon Valley. We really love developer tools, building for developers because we feel the pain. In China, there’s perhaps more B2C focus, which actually works to their benefit in terms of finding good use cases.
Rohit Krishnan: What is the previous large software success story from China that took over the world? There’s TikTok, and that’s essentially it. WeChat is amazing, but nobody uses it outside China — maybe in parts of Southeast Asia. Banking software emerged, but nobody really adopted it. Alibaba Marketplaces exist, but they haven’t permeated the West in any meaningful sense.
This might be an unconventional statement, but AI is one of those domains where you can build amazing AI agents using existing models from anywhere in the world. I’m glad we’re starting to see that happening.
Jordan Schneider: It was remarkable how this company, with both its browser and its first product Monica — a ChatGPT-like search browser add-on — targeted foreign users first. That’s notable because running Claude is illegal in China, which makes development difficult.
Reading interviews with the CEO over the weekend, he stated essentially: We’re not really trying to take on the big labs, but we think there’s an opportunity and a big market here. It was somewhat sad reading when he alluded to the politics of AI: “I've come to understand that many things are beyond your control. You should focus on doing well with the things you can control. There are truly too many things beyond our control, like geopolitics. You simply can't control it—you can only treat it as an input, but you can't control it.”
Frankly, I don’t think Chinese AI agents will have much longevity in the US market without hitting some severe regulatory headwinds. However, their skill at playing the global influencer marketing game to generate this hype cycle reflects a real fluency in digital marketing. The fact that they could play this game better than any Western agent competitor — except for Devin, which tried but faced its own challenges — is remarkable. There hasn’t been another major attempt at this over the past year and a half.
Dean Ball: I would go even further. When the first Devin demos appeared, people exclaimed, “Look how cool this thing is!” Then the bubble burst when people realized it was GPT-4 with prompt engineering and scaffolding.
The Western AGI obsession makes us want to conceptualize this as one godlike model that can do everything, and we implicitly dismiss product engineering and practical applications. You see that reflected in public policy, which is obsessed with big models, giant data centers, and similar infrastructure. Those are the only things we seem to take seriously and value.
I’m a deep learning optimist — I’m not going to tell you AGI doesn’t exist or is a Gary Marcus fiction. I’m not in that camp at all. But the AGI obsession has developed into something that feels like a perversion, distracting us from opportunities lying right in front of us.
I’m not necessarily saying Manus represents that opportunity, but there are thousands of possibilities where cleverly stringing together different AI products and modalities could yield interesting results. We just don’t see much of that happening. A year ago, I was more inclined to say, “Well, it takes time,” but a year later, I find myself less willing to make that excuse.
Structural Factors Driving the AI Product Overhang: Why Big Labs Don’t Do Product
Rohit Krishnan: Shawn, what is this? Are the VCs dumb? Are the founders dumb? Are there actually not pennies to be picked up off the ground?
Shawn Wang: There are, and the VCs have woken up to it. I started writing about the rise of the AI engineer two years ago, and now there are VPs of AI engineering at Bloomberg, BlackRock, and Morgan Stanley — they just spoke at my conference last week.
People were very dismissive of the GPT wrapper, viewing it as just a thin layer over the LLM. Now the perception has almost flipped, where the model is the commodity and everything else on top of it is the main value and moat of the product. This is why I started talking about AI engineering, and I think it’ll be a growing job title. It’s what we orient my conference and podcast around.
It’s music to my ears. I’ve been saying this for a while now, and the VCs have caught up. It’s just harder to fund because you can’t just say, “Here’s the pedigree of the 10 researchers I have. Give us $300 million.” Now you have to actually look at the apps and see if they’re well-engineered and fit the problem they’re trying to solve, whether B2C or B2B. That’s much more difficult than throwing money and GPUs at talented researchers and letting them go for it. That approach caused Inflection AI, Stability AI, and other mid-tier startups to burn around $100 million each.
Rohit Krishnan: That’s back to the SaaS era in some sense. You suddenly find a new vertical niche where you can build something, spend time and effort, learn about the specific problem you’re solving — not just intelligence but something more targeted, B2C or B2B. Then you have to tackle it and solve it.
Shawn Wang: The interesting thing about this SaaS transition is you’re charging on value and not on cost, and the margins between those approaches are enormous. Many of us in Silicon Valley realize that if you develop your own models, the next one that comes out is probably open source from China and better than yours. So where’s the value in that?
Everything’s being competed down to cost. Anthropic offers Claude at cost. OpenAI has a small margin, but every other GPU provider serving open source models is just providing at cost because they’re trying to capture market share with VC money. Nobody’s making a margin here.
You contrast the $200 versus the potentially $2,000 or $20,000 a month agents you can offer, because you’re competing against human labor and human thinking time, which we are all limited by. The economics start to really work out. You could start charging for your output instead of charging for your cost of goods sold. That is fundamentally a better business.
Rohit Krishnan: Speak for yourself, Shawn.
Shawn Wang: The fact that you could just start charging for your output instead of charging for your cost of goods sold is fundamentally a better business.
Rohit Krishnan: You wrote a very cool thing about Google’s awkward struggles to make products that people use. What is stopping the big model makers from starting to do things they can charge value-based pricing for? Is it just that they don’t need to and have their hands full making AGI?
It does seem that just selling tokens isn’t going to make you money in the long run. It’s funny because if you’re one of the large labs, if you’re Sam or Dario, you don’t particularly care about that since you already have so much money coming your way. Anthropic just raised $60 billion, OpenAI is valued at $300 billion. These are astronomical figures.
We’ve normalized these numbers in conversation, but they’re absurd by any stretch of imagination. $300 billion is bigger than Salesforce. It’s insane to think about for a company. Why are they getting that money? Because they want to build AGI. Why do they think they can build AGI? Partly because they’re true believers, partly because they have the best research talent in the world who wants to build AGI.
What happens if you tell that research talent that they’ll be working on building agents for awhile? Many of them quit. Arguably, many did. In a weird way, it’s only in larger places like Google where you can potentially have a large enough contingent of people try some unusual approaches and build cool stuff.
They did create interesting products — NotebookLM is actually really interesting. It was a cool new product, new modality, new way of interacting with information. I am surprised that we didn’t see more of it. In typical Google fashion, it just kind of disappeared after a while. They have Colab, which is an interesting product that’s languishing in a corner somewhere.
Everything Google does involves creating a very interesting first product and then slowly killing it by cutting off the oxygen supply over the next five years. For somebody to care deeply about building a product here, it has to start right at the top. It has to come from a mission, because the argument against building a product — the engineers saying, “Just wait a year and everything will get solved" — is really seductive.
Safety, Liability, and Regulation
Shawn Wang: You really need somebody who has a Jobsian level of ability to push back and say, “I don’t care what you guys think. We need to actually build something that really works here.” That’s not a muscle that any of these companies have because none of them have built products. Arguably, the thing that kicked it all off, ChatGPT, was built as a research preview. What we are doing is all being okay with playing around with research previews that consistently sneak their toe in and pretend they’re a bit of a product, but they’re not really.
Rohit Krishnan: Let’s fast-forward to the near future when agents can do more economically useful things than book you a train ticket. Should we start with the safety angle? It’s wild if I’m going to let something exist as me on the internet or in my workplace and I’m responsible for it. Maybe Manus is responsible? Maybe OpeningEye’s responsible? Maybe the AI engineer who goes to Shawn’s conferences is responsible?
This is a very weird world where it’s not just Jordan Schneider as an AI-enhanced worker using chatbots, but Jordan Schneider letting go a little bit and having these automated minions exist under my aegis but also not.
Dean Ball: I haven’t checked their website thoroughly, but I would be very surprised if Manus or the company that built it has a safety and security framework, a responsible scaling policy, or has commented on the EU code of practice.
Rohit Krishnan: I actually looked for this. I could not find one thing that the CEO has said in any relation to any safety discussion or question.
Dean Ball: This thing doesn’t have any guardrails. I don’t think it’s a consideration for them. In some sense, that’s probably part of what makes this better than Operator or Claude computer use, because Anthropic and OpenAI have both legitimate business incentives and internal stakeholders who won’t let the company ship things with no guardrails.
There’s reputational risk. If OpenAI had released something like Operator with zero guardrails, you’d be looking at state attorneys general investigating you, and the FTC and others coming after you, just as they did with ChatGPT. The tech industry is pretty risk-averse on things like this because it’s an inherently risky endeavor.
Those are market incentives, because you shouldn’t be incentivized as a consumer or business user to throw agents into the wild who do things for you and potentially cause problems. There should be some liability for that. You should be incentivized not to do such things, and companies should be incentivized not to release such things.
I’ve been thinking about liability issues in the last few months and have concluded that the court system is going to really struggle. If something happened with Manus, there’s the user who prompted it, multiple LLMs behind the scenes, and a Chinese company that is almost certainly not subject to a legally cognizable claim, unless you want to go to court in Beijing. How is the American tort liability system going to figure this out? I’m skeptical it will do a very good job.
But no liability is a moral failing, too. As the cost of cognitive labor declines, one of the only things left with economic value is trust, pricing risk, and similar concepts. I wonder if frontier AI companies will slowly converge to being more like insurance companies or financial services companies. Those industries are based on trust, pricing of risk, and allocation of responsibility for harm that occurs from realized risks. That feels like what’s economically valuable here, certainly not selling marginal tokens.
Shawn Wang: There’s one proof point that maybe agrees on some level: we’ll never get the O3 API because OpenAI is choosing to release products instead of APIs. That makes sense if you believe your APIs are valuable — you stop giving them to everyone else. It also stops the Manuses of the world in their tracks, because they can no longer use those APIs.
In broadening this general safety discussion, this is just an argument for American AI accelerationism. The simple fact is, if you are more safe and stop yourself from doing anything, then China will do it first, and you’re behind. It’s better to be ahead and in control of the narrative, build in the safeguards at the LLM layer with the post-training that you do, and try to lead from the front instead of the back.
Rohit Krishnan: I have a more contrarian view. Even framing this in terms of safety is incorrect. What are we talking about today? The Manus of today, Operator of today — these aren’t safety concerns. They’re engineering concerns, misuse concerns. We’re using the 2023 version of AI safety, which seeps into every part of the “anything a model can do can be unsafe” conversation, and that distorts how we discuss what these products do.
As Dean said, the liability issue is important once these tools start getting used inside companies. If someone at Pfizer uses Manus to figure something out and creates a wrong drug, there are liability issues. If someone at Cloudflare uses Manus to fix a bug and creates an outage, there are clear questions about where responsibility sits.
But we’re still at the point of making these things work properly in the first place. Think about our example — testing if it can book an Amtrak ticket. We’re not yet at the point where AI agents are so incredibly amazing that we have to restrict them before they engage. I’d like to see them work properly before we leash them.
That doesn’t mean we shouldn’t have a parallel track thinking about liability issues. But these will be hard-won battles that push the frontier forward one issue at a time, rather than “We’ve figured it out for everything from searching medical information to booking tickets to writing open source code or malware."
One problematic outcome of these discussions in recent years is that we’ve conflated all these issues into one, and they’re not the same. I look at Manus and think, “Good. I’m glad somebody without a responsible scaling policy is showing us what can be done,” because there’s no inherent problem with giving something a browser. Yes, there can be prompt injection attacks — that’s new and we need to solve it, but we can’t figure it out without anybody actually doing anything. It’s a chicken and egg issue.
Dean Ball: If you’re a dentist trying to use AI to automate business processes within your dental practice, then the fact that OpenAI has a responsible scaling policy about biological weapons risk evaluation isn’t that important for you. But perhaps more problematically, OpenAI’s model specification says, “Follow the law.” Okay, gotcha.
My view is that we have to almost entirely reject the tort liability system for this because it’s too complicated an issue. This is the kind of situation where transacting parties need to come to agreement about what makes sense in these particular contexts and let contracts do their thing. The courts won’t adjudicate this on a case-by-case basis in any effective way.
The risk of accelerationism is that you accelerate without proper safeguards. Noam Shazeer left Google to accelerate and founded Character.AI. What happened? Character.AI said problematic things to children, made sexual advances to children, and a kid killed himself. I don’t know if you could say Character.AI is responsible for that child’s suicide, but he was talking to the chatbot when he killed himself. That’s a tort case — Tristan Harris is funding it in the State of Florida, with a sympathetic jury and judge.
What’s Character.AI now? It’s a husk. Noam Shazeer’s gone, back at Google, and the company is likely to be picked apart in tort litigation, with other cases against them too.
If you accelerate without figuring this out, something very bad could happen. As they say, bad facts make bad law. Maybe it’s not unambiguously the AI model’s fault, but if there’s a really nasty set of facts, you could get adverse judgments in American courts. The common law is path-dependent, so you could end up with a very bad outcome quickly.
I’m enthusiastic about accelerating adoption and diffusion — Manus is very much a diffusion story. But if we don’t, in parallel, work on risk assignment (not catastrophic risk safety, but determining who is responsible when things go wrong), we could end up in a bad situation rapidly.
Legal Frameworks and Innovation Timelines
Shawn Wang: Do you think it’s primarily financial infrastructure that is needed, like your model of AI companies as insurance companies?
Dean Ball: Legal and financial, yes. What you basically need is a contracting mechanism that is perhaps AI-enabled — AI-negotiated contracts, perhaps AI-adjudicated contracts so you don’t have to deal with the expense of the normal court system. Once you have contracts and liabilities on the balance sheet, you’re in derivatives territory.
It’s a New York problem, not a San Francisco problem at that point. This is certainly an AGI-pilled idea. I wouldn’t do this with Claude 3.7, even though I think it’s great, but I think we could get there in the next couple of years when models are capable of doing things like this.
This is just one approach, certainly not the only one, and it’s a nascent idea for me. But this could be where the money actually is — pricing risk and transforming risk is something America is much better at than China. We’re fantastic at that.
It’s weird because many of my Republican friends in DC hate that fact. They view finance as decadent, as does Chairman Xi. But there might be trillions of dollars of wealth to be created here.
Shawn Wang: Any financial asset is based on the laws it’s grounded in, and I think the laws have to be figured out here. There’s a bit of a chicken-and-egg situation with that.
Dean Ball: Yes, but if you have contracts, contracts would form a substantial part of the law.
Shawn Wang: They still need to be litigated. One measure I’d be interested to plot is AGI timelines versus legislative timelines. We’re accelerating in AI progress and decelerating in law and Supreme Court resolutions of cases. Our legal infrastructure needs to keep up with AI progress, or we’re in serious trouble.
Dean Ball: I completely agree. That’s the problem I try to get my head around all the time.
Jordan Schneider: We saw the EU just try and completely fall on their face, which was not a good first effort for democracies.
Dean Ball: The problem is you don’t want to create a statute prematurely — a statute with a bunch of technical assumptions embedded in it prematurely. It’s a very narrowly targeted thing. For me, this is all clicking into place, and I think if we got this done in the next two years, we’d be fine.
Jordan Schneider: The future of agents in China is going to be really interesting. There was an argument two years ago that LLMs would have a hard time gaining traction in China because the government would worry about aligning them to avoid anti-party statements. But this is basically a solved problem.
I’m curious about your perspectives on the technical challenge — not just at a legal level of assigning blame, but at a product and operational level of building things that governments and large companies will be comfortable with. Is this just a matter of time? Is there anything fundamentally difficult requiring major breakthroughs? Once we have the technology to make Operator and Manus do really good things, will they be controllable as well?
Dean Ball: I’d be curious if you’d correct this assumption if I’m wrong, Jordan, but my impression of China is that it’s actually a somewhat more ice-cold libertarian country when it comes to liability issues, where there’s a greater “developing country” or “these things happen” mentality.
Jordan Schneider: Yes, until bad things happen, and then your company gets shut down.
Dean Ball: It’s more of a binary outcome.
AI and the Future of Work
Jordan Schneider: Let me take this in a different direction. JD Vance at Paris said, “We refuse to view AI as a purely disruptive technology that will inevitably automate away our labor force. We believe and we’ll fight for policies that ensure AI is going to make our workers more productive. We want AI to be supplementing, not replacing work done by Americans.”
Shawn Wang: This is something AI engineers worry about a lot. A surprising number of them are actually worried for their own jobs, which is very interesting.
The main question is whether you have a growth mindset or a fixed mindset view of the world — whether you believe human desires tend to expand over time. Whenever we reach a certain bar, we immediately move that goalpost one football field away. The idea is that, yes, AI will take away jobs that exist today, but we will create the jobs of tomorrow, and ideally those are the jobs we want to do more of anyway.
Rohit Krishnan: I agree. To a large extent, that sentiment is the most normal politician statement in the world — technological growth is great and will continue making everyone’s lives better. It’s the same thing people have said for a very long time.
The difference here is that there’s at least a contingent of people who look at that and say, “No, this time it’s different.” You might say it’s not disruptive, but it could be massively disruptive in a short period of time to a large segment of society. It’s not just agriculture getting mechanized, but potentially all white-collar jobs.
When I’ve examined this issue, I don’t think massive disruption will happen immediately. The technological, regulatory, and sociological barriers are large enough that we won’t all be unemployed in five years. There are enough things to do. As Shawn mentioned, we’ll have to address the inevitable complications of regulatory frameworks before these technologies can be deployed everywhere.
When I did some rough calculations, we’ll still be bottlenecked by chips and energy in 10 years, which will prevent us from replacing all labor with AGI or AI agents. Does that mean there will be no disruption? Absolutely not. I fully expect disruption.
We already have AI that can plausibly replace large chunks of specific white-collar tasks that I do, you do, legislators do, or Supreme Court justices do. Pick your poison — we could probably replace a chunk of these roles with Claude 3.7 and get better results. We’re already in that world, but complete transformation won’t happen immediately.
JD Vance has to toe the party line: anti-AI safety, pro-acceleration, technological optimism all the way. I support that approach, but I’m not parsing his statement with any deeper meaning than that.
Dean Ball: I find myself more worried about slow diffusion due to multiple factors. The bottlenecks are regulatory, but there are many other bottlenecks as well. I’m much more concerned that diffusion and actual creative use will be slowed. I worry about the uses of LLMs that no one has ever thought of, and I’m concerned that no one will ever think of them. That’s probably the bigger issue we should be addressing through public policy.
In the longer term — and in AI time, that’s about three years — I do think there’s a possibility that some elite human capital might get automated in different ways. Political instability tends to emerge when you have an overproduction of elites in a society. We already have that problem. We’ve already significantly overproduced elites in America, and I worry that will get worse. When you combine that with other political problems America faces, you could have a tipping point phenomenon.
I wouldn’t dismiss it entirely as a risk, but my default assumption would be that the risk is actually on the other side — not diffusing fast enough.
Rohit Krishnan: I have a thesis that I sometimes hold that markets that become extremely liquid end up with polarized outcomes. We’ve generally seen that with capital markets — globalization has meant some companies get extremely large while the middle gets decimated, which is why most gains come from the Magnificent Seven.
We could easily see something similar happen in labor markets. We already see flashes of it. Engineering salaries have a somewhat bimodal distribution. Lawyers experience it too. Once AI enters the picture, we might have a vastly more liquid labor market than ever expected. This sounds nice, except it results in a power law distribution. Polarization is difficult to address in domains we don’t know how to handle cleanly or where we can’t easily establish minimum thresholds.
Jordan Schneider: Any closing thoughts?
Shawn Wang:[In true professional podcaster fashion…] I don’t know if we’ve answered the question that you’re likely to put in the title of your episode: “Is this a second DeepSeek moment?” For what it’s worth, my answer is no.
Dean Ball: I agree. In some sense, I think it’s actually more interesting than DeepSeek. And in another sense, it’s certainly not as impressive of a technical achievement as DeepSeek.
Rohit Krishnan: I’ll argue for yes, because I think DeepSeek was a DeepSeek moment for core research talent. Manus is closer to DeepSeek for product. I’m glad we’re pushing a second boundary as opposed to pushing the same boundary.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
Just as the buzz around DeepSeek was beginning to fade, Chinese AI has made waves again with the AI agent “Manus,” launched on March 6th, 2025.Today, we’re here to unpack the Manus launch, explore the business model of Manus’ parent company, and offer a glimpse into the mind of Xiao Hong 肖弘, the founder behind China’s latest viral AI product.
Manus claims to be the world’s first general-purpose AI agent. It ostensibly outperforms OpenAI’s ChatGPT Deep Research on the General AI Assistants (GAIA) benchmark. Currently in beta testing, access is restricted to those with invitation codes, which are reportedly being listed second-hand for 50,000-100,000 RMB (whether anyone actually paid that much is another question). Users report impressive performance in basic tasks, for instance rebooking airline flights, beyond what Anthropic’s Computer Use and OpenAI’s Operator have thus far provided to users. The product is also experiencing slowing response times, hinting that Monica.ai may be struggling to scale up compute to meet skyrocketing demand.
Manus started in 2022 as an AI-powered browser plugin, backed by ZhenFund (真格基金). In 2023, the company secured Series A funding led by Tencent (腾讯) and Sequoia Capital China (红杉资本中国). What began as a simple “ChatGPT for Google” browser plugin has since evolved into a full-fledged AI agent.
Monica, the company that developed Manus, operates from Wuhan, rather than from China's major tech hubs like Beijing or Shanghai. In early 2024, ByteDance attempted to acquire Monica for $30 million, but founder Xiao Hong (肖弘) turned down the offer. ByteDance’s plan was to absorb Monica’s team and technology into its Doubao AI ecosystem, a move that would have diluted Monica’s distinct market position. Instead, Monica closed a new funding round at the end of 2024, reaching an estimated valuation of nearly $100 million.
The exact AI models powering Manus remain unclear. The company claims to use multiple models for different tasks. Notably, when prompted to reveal its own system files, Manus reveals it may be powered by Anthropic’s Claude models — which would make operating in China illegal. This probably explains why Monica’s website appears to be blocked in China.
Edit: confirmed by co-founder.
Anyway, the fact that Manus appears to disclose more than it should hints at broader potential security vulnerabilities.
Who is behind Manus?
Founder & CEO, Xiao Hong (肖弘), is a serial entrepreneur and a graduate of Wuhan’s Huazhong University of Science and Technology (华中科技大学). He first made his mark by building WeChat-related tools as a student, admitting that while his “academic performance was quite poor,” he partnered with more technical classmates to build tools. In 2015, he launched Nightingale Technology (夜莺科技) and created Yiban Assistant (壹伴助手), a WeChat management tool that secured early backing from ZhenFund (真格基金).
By 2019, Xiao saw a bigger opportunity in enterprise WeChat tools and developed Weiban Assistant (微伴助手). His timing was perfect—when rival WeTool (微商工具WeTool) was shut down in 2020, Weiban became the go-to alternative, attracting investment offers from Sequoia Capital China (红杉资本中国) and Youzan (有赞). Eventually, Minglue Technology (明略科技) acquired Weiban, marking Xiao’s first major financial success.
Sensing the potential of large AI models, Xiao left Minglue in 2022 to create Monica.ai, originally designed as a “ChatGPT for Google” browser plugin.
Co-founder & Chief Scientist, Ji Yichao (季逸超) dropped out of high school at 17 to develop Mammoth Browser (猛犸浏览器). His talent caught the eye of Sequoia Capital China’s Zhou Kui (周逵), who introduced him to investor Xu Xiaoping (徐小平). Xu invested 1.5 million yuan, giving Ji complete creative freedom. Recognizing the large potential of LLMs, Ji joined Xiao Hong to start Monica in late 2022.
Interview Quotes
Unlike DeepSeek’s media-shy Liang Wenfeng, Xiao Hong has done a ton of press. Below are selected translations from several in-depth interviews with Monica’s founder and CEO, Xiao Hong, offering insights into his vision, strategy, and the future of AI agents.
The vibe of Xiao Hong’s interviews is distinct from the AGI-driven idealism blended with national pride we’ve seen from the founders of DeepSeek and Unitree. Xiao is pragmatic and focused on profitability rather than research. A newly published three-hour podcast with Xiao opens with offering this piece of advice:
“I remember there was a Northeastern Chinese restaurant near my university. I made enough money to treat my tech club friends to dinner there every day. Here’s a tip for the audience: if you’re in college, take your most talented classmates out for meals as often as you can. If you wait until after graduation to recruit them for your startup, you’ll have to treat them to Michelin-starred restaurants instead.”
In another interview from January 2024, Xiao openly admits that he didn’t initially believe in AI’s potential, and “remained cautious” despite the hype surrounding GPT-3.5 in the fall of 2022. He describes coming to two conclusions about AI investment, which eventually led him to focus on AI products as opposed to chasing AGI with foundational model research:
"First, I wouldn’t consider working on big models without sufficient business scale. Second, I believe that in China, big model services will eventually integrate fully with cloud computing. I’ve discussed this with our CTO and believe that cloud computing companies will provide customized deployment services, so we don’t need to dive into that ourselves."
…
"I focused more on what big models could do, and what kind of applications I could build with them. In the beginning, many people were financing based on concepts, but by the second half of the year, both domestic and international, there was much less of that. Everyone was returning to business rationality, focusing on finding PMF (Product-Market Fit). By February of 2023, I had a conversation with an investor focused on big models, and no matter how I asked, they refused to talk about products. They weren’t discussing technology or plans. By March, the product’s valuation plummeted. People realized that simply building a single application based on big models might not work, and that’s when the consensus started forming: either focus on technological breakthroughs or work on relatively closed-loop application scenarios."
…
"In March and April of 2023, the fastest-growing product outside of ChatGPT globally was Poe. It was essentially a shell around a big model, and I told investors that if you can perfect the shell, that’s still a big deal. So we decided to do it too, and instead of resisting the demand, we decided to embrace it. In the first half of 2023, Monica integrated all the major models because that’s what the users wanted, and we started by doing that, figuring out how to find more use cases step by step."
Monica’s business model focuses on catering to the overseas market, which likely explains why their website is devoid of any reference to being based in China. Besides English, Monica’s website has dedicated versions in traditional and simplified Chinese, as well as Russian, Ukrainian, Bahasa Indonesian, Persian, Arabic, Thai, Vietnamese, Hindi, Japanese, Korean, and a slew of European languages.
In Xiao’s words, “We chose to target the overseas ToC market because I felt it was a larger, more commercially viable market. The domestic market seemed a bit more challenging.” Their focus shows: in contrast to DeepSeek’s very low key model launches, Manus’ launch came with a whole sophisticated press push like one you would see out of a YC startup, complete with a very well-produced English-language launch video and early access for select YouTubers and twitter influencers.
International expansion comes with its own difficulties, but Xiao believes those challenges made Monica stronger as a company. He’s recently argued that China would benefit from having more firms look abroad:
Xiao Hong: I think we are still in a great era with many opportunities…. First, it's the AI era. Second, I think we are also in a great era of globalization. I'm not a geopolitical expert, but it seems like every country has its own problems — internally, everyone has their own issues. So overall, the world is becoming more conservative and more isolationist, right? But at the same time, no one wants others to be isolationist; they only want to be isolationist themselves. So, everyone hopes that their own entrepreneurs will think more globally.
I believe China’s entrepreneurs of today should be more aggressive in globalizing. If we see overseas markets as better opportunities, it’s not just about market-driven decisions — we should step into international markets to gain experience. We need to participate in global competition, rather than just competing in the markets we are familiar with.
By the way, this process requires a lot of things. When I started this company, none of our founders had lived abroad for an extended period. Everyone’s English proficiency peaked in high school and declined in college! [
I once joked that if, at the same time, there was another founder who had lived in the U.S. and was placed next to me, I would have chosen to work with that founder myself. But this shouldn’t be the way we compare things — it should be about doing our own thing. Secondly, I had a simple belief at the time: the global market is much bigger, and the market itself will provide the tuition fees for founders to learn. (Laughter)
Besides the AI era, another crucial topic is that we are now thinking about things with a globalized mindset.
Unsurprisingly, this business model also relies on collecting vast amounts of user data. Monica’s free Chrome extension requests expansive access to browser data, including permission to log keystrokes, and Manus “crawls” devices to make suggestions. Xiao is betting that widespread adoption of these products will unlock a treasure trove of monetizable insights.
“The data we collect through our browser plugin is critical. Even though this might not guarantee success, it’s a step in the right direction. The private data we gather, along with contextual information, will help differentiate us from the competition. This is one of the key assets we need to grow.”
Xiao is explicitly describing an intent to build an incumbent advantage on a foundation of user data, and TikTok demonstrates how effective that strategy can be. Reliance on eventual mass adoption could partially explain the high-publicity invite-only launch strategy for Manus (although limited access to compute is also certainly a factor).
That said, he is aware that the politics exist and could get in the way of a Chinese-owned AI agent gaining widespread adoption abroad. He spoke about it in a recent podcast alluding to NeZha 2.
I've come to understand that many things are beyond your control. You should focus on doing well with the things you can control. There are truly too many things beyond our control, like geopolitics. You simply can't control it—you can only treat it as an input, but you can't control it.
I recently asked DeepSeek to explain three terms 贪 (greed), 嗔 (hatred), and 痴 (ignorance) [the ‘three poisons’ of Buddishm recently spotighted in the truly excellent animated movie NeZha 2]. It explained it very well: greed is attachment to favorable circumstances; anger is dissatisfaction with adverse circumstances; and ignorance is not understanding the truth of the world. The "truth of the world" is very profound, so I won't discuss that. But greed and anger are problems many people encounter, as are attatchment to favorable circumstances and dissatisfaction with unfavorable ones.
This business-minded pragmatism shines through in Xiao’s vision for the future — instead of techno-optimist visions of AI-powered drug discoveries or a moon colony staffed by robots, he imagines a world where humanity can return to a glorious past:
“I think that the white-collar lifestyle may be a detour for mankind. If you look at it in terms of a curve or over a longer period, say thousands of years, or even the ten-thousand-year span of human history — it's actually quite rare for people to sit in one place and engage in intense mental work without much physical activity. This is probably only a phenomenon of the past hundred years.
For a longer time in history, maintaining physical health and developing spiritual civilization have gone hand in hand. In ancient times, people also needed spiritual and cultural development, but that involved physical labor as well, which helped strengthen their bodies.
In the past hundred years, however, issues like diabetes and high blood pressure have become widespread because people work in this sedentary way. If we look at humanity as a whole, sitting and working for eight or more hours a day is an anomaly.
If AI can take over these tasks, then people can work fewer hours and go back to living more like they did in the past — focusing more on spiritual and cultural enrichment while also taking better care of their physical health.”
To close, here’s a quote from Xiao about how it feels to live through history:
Xiao Hong: From the time I was born in the 90s until now… there have been significant shifts, from PCs to mobile, then the semiconductor industry, which has been booming behind the scenes, the rise of the internet, and now artificial intelligence. I feel like these opportunities are emerging very intensively. When I watched The Godfather, I realized that if I had lived in that era — it was also a time of change — but if you lived in certain periods, you might not have witnessed such rapid technological progress. Sometimes, when we read history books or ancient texts, it feels like things barely changed, which I think would be a little frustrating!
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
TLDR: U.S. export controls targeting China’s AI capabilities focus primarily on limiting training hardware but overlook the growing importance of inference compute as a key driver of AI innovation. Current restrictions don’t effectively limit China’s access to inference-capable hardware (such as NVIDIA’s H20) and don’t account for China’s strong inference efficiency. While China’s fragmented computing infrastructure has historically been a disadvantage, the shift towards inference-heavy AI paradigms positions their compute ecosystem to be more utilized and valuable. As reasoning models, agentic AI, and automated AI research elevate the role of inference to advancing AI capabilities, the US should urgently strengthen export controls to hinder China’s inference capacity and develop a coherent open-source AI strategy to maintain competitive advantage.
The Export Control Status Quo is Broken
The global AI competition is unfolding along two critical axes: innovation — the development of advanced AI capabilities — and diffusion — deploying and scaling those capabilities. The United States has prioritized outpacing China in AI innovation by focusing on pre-training as the main driver of progress. However, a new paradigm is emerging where inference, not just training, is becoming central to advancing AI capabilities.
This shift has significant implications for U.S. AI policy. Current export controls aim to limit China’s ability to train frontier AI models by restricting access to advanced chips, based on the belief that scaling pre-training is the primary driver of AI progress. By limiting China’s access to compute resources, these controls aimed to slow its AI development.
Yet, these same controls are far less effective at restricting China’s inference capabilities — exposing a critical gap in U.S. strategy. As inference becomes more central to AI innovation, current policies are increasingly misaligned with the realities of AI development. To effectively counter China’s growing inference capabilities, the U.S. must strengthen its export controls.
Inference Compute is a Key Driver of AI Innovation
The AI landscape is evolving beyond the scaling pre-training paradigm that dominated recent years. Emerging solutions are shifting innovation toward a paradigm where inference compute — not just training compute — has become a critical driver of AI progress.
Three interconnected trends are driving the link between inference and AI capabilities:
Reasoning Models
These models require significantly more inference than traditional LLMs, leveraging test-time scaling laws that suggest a link between amount of inference compute and model performance. Inference demand is further driven by a feedback loop that accelerates AI capabilities: reasoning models generate high-quality synthetic data, which enhances base models via supervised fine-tuning (SFT). These stronger models can be adapted into stronger reasoning systems, creating even better synthetic data and fueling continuous capability gains.
Agentic AI
AI agents — systems capable of taking autonomous actions in complex environments to pursue goals — are often powered by reasoning models, which drives up inference demand. Many agents have access to external tools and environments such as code execution environments, databases, and web search, which enhance their capabilities by enabling them to retrieve information, plan, and interact with digital and physical environments.
Some agents continuously learn by interacting with their environment via reinforcement learning. Unlike standard language models that handle one-off queries, agentic AI systems require persistent inference as they continuously interact with external environments, adapt to new information, and make complex, multi-step decisions in real time — significantly increasing overall inference requirements.
Automated AI Research
Automated AI researchers can design new architectures, improve training methods, run experiments, and iterate on findings. Scaling in this paradigm requires both inference compute to power research agents and training compute to execute their proposed experiments. Greater inference capacity allows more of these systems to operate in parallel, expanding both the breadth and depth of AI exploration. This, in turn, enlarges the search space they can navigate and directly increases the rate of AI innovation.
Greater inference also enhances research agents through iterative reasoning, self-play debates, and automated evaluation — capabilities already demonstrated in AI-driven scientific discovery. As these automated systems achieve early breakthroughs, they become better at identifying promising research directions and architectural improvements, potentially setting off a compounding cycle of progress. Thus, even small initial advantages in inference capacity can compound, leading to a significant, potentially decisive, lead in AI capabilities.
In an era of reasoning models, agents, and automated AI research, inference capacity is not just an enabler — it is a primary determinant of the speed and trajectory of AI innovation. This shift has significant implications for the U.S.-China AI competition and underscores the need for stronger U.S. export controls.
China’s Inference Capacity is Key
Current U.S. export controls aim to restrict China’s ability to train frontier AI models but overlook the growing importance of inference and China’s capacity to scale it. As AI development shifts towards inference, China’s position strengthens considerably due to three key factors:
Steady access to inference-viable GPUs
Leading inference efficiency
Compute ecosystem being better suited for inference rather than pre-training
Access to Inference Hardware: The H20 Loophole
Despite U.S. export controls restricting access to cutting-edge AI chips like the H100 and H800, China maintains strong access to inference-capable hardware through several avenues — most notably through Nvidia's H20 GPU.
The H20 represents a significant gap in current export restrictions. Specifically designed to comply with export controls and serve the Chinese market, the H20 is actually superior to the H100 for particular inference workloads. The H20 outperforms the H100 for inference workloads due to its superior memory capacity and bandwidth. It delivers 20% higher peak tokens per second and 25% lower token-to-token latency at low batch sizes—key advantages given that inference performance is driven more by memory bandwidth and batch efficiency than by raw computational power. With 96GB of HBM3 memory and 4.0TB/s memory bandwidth, compared to the H100’s 80GB and 3.4TB/s, the H20 is highly viable for inference, making it a significant gap in current export restrictions.
Figure 3: GPUs restricted under iterations of U.S. export controls. Source: SemiAnalysis x Lennart Heim
China has been importing large sums of the H20. SemiAnalysis estimates that in 2024 alone, NVIDIA produced over 1 million H20s, most of which likely went to China. Additionally, orders by Chinese companies, including ByteDance and Tencent, for the H20 have spiked following DeepSeek’s model releases.
Access to Inference Hardware: Trailing-Edge GPUs
Trailing-edge GPUs remain surprisingly effective for inference workloads. China retains strong access to trailing-edge GPUs due to largestockpiles of the A100, A800, and H800 in 2022 and 2023. Additionally, Chinese firms, including Huawei, Alibaba and Biren, have also developed indigenous chips. The viability of trailing-edge GPUs for inference suggests that China’s inference capacity is stronger than their volumes of cutting-edge GPUs may suggest.
The effectiveness of older GPUs for inference stems from fundamental differences between inference and training workloads:
Long-Context Inference is Memory-Bound, Not Compute-Bound
Unlike training, inference only runs forward passes, avoiding computationally intensive processes like backpropagation and gradient updates. As a result, inference is significantly less compute-intensive than training.
The real constraint for inference is memory. Inference, particularly long context inference, is currently memory-bound rather than compute-bound due to several factors:
Model Weights & Key-Value (KV) Cache: For transformer-based models, inference requires storing both the model parameters and a key-value (KV) cache. The KV cache stores the past tokens' key-value pairs, allowing the model to retain context and coherence, and grows linearly with the context length. While compute resources are only required to process each newly generated token, memory usage continuously increases as new key-value pairs for each transformer layer are stored in the cache with every additional token generated. Consequently, total memory consumption rises steadily as the context expands, in contrast to compute needs, which remain stable and do not accumulate in the same manner. As a result, inference often becomes memory-constrained before it becomes compute-constrained, particularly for long-context tasks, where the KV cache can exceed the model weights in size.
Autoregressive Bottleneck: Input tokens can be processed in parallel, leveraging the full sequence since it’s known upfront. However, output tokens are generated sequentially, with each new token depending on all the previously generated tokens. This creates a bottleneck during output generation:
Full KV Cache Access: Each generated output token requires accessing the entire KV cache.
Memory Bandwidth Limitation: On long sequences, this repeated full KV cache access for every output token creates a memory bandwidth bottleneck (data transfer rate between memory and processor), which becomes the primary limiting factor.
Constrained Batch Sizes: The size of the KV cache directly limits batch size during output generation. Longer sequences consume more GPU memory, reducing space for batching multiple sequences. This forces smaller batch sizes–the amount of independent user queries that can be processed in parallel–which reduces GPU utilization and restricts inference throughput.
This memory constraint becomes evident when examining FLOP utilization rates. During inference operations, GPUs typically achieve only about 10% FLOP utilization when generating tokens, compared to 30-50% during training. This underutilization occurs because GPUs spend much of their time retrieving and managing the KV cache rather than performing actual computations. The inefficiency grows even more pronounced with newer, more compute-dense chips, where increasingly powerful processing cores sit idle waiting for data to arrive from memory.
The Memory Wall
This inference bottleneck reflects a broader structural limitation in computing hardware. While GPU compute performance has grown exponentially (approximately 3.0x every 2 years), memory bandwidth and capacity have improved at a much slower rate (around 1.6x every 2 years). This growing gap creates a “memory wall” where performance is constrained not by processing speed but by how quickly and how much data the GPU can store and access.
Fig 1: Memory, in green, has scaled at a lower rate (1.6x/2yrs) compared to computational performance, in black (3.0x/2yrs). Source: Gholani, Amir, et.al. (2024), AI and Memory Wall.
This memory-bound nature of inference has significant implications for hardware viability. While newer GPUs offer exponential improvements in raw computational power (measured in FLOPs), they provide more limited gains in memory capacity and bandwidth — the true bottlenecks for inference workloads.
As a result, inference workloads often cannot fully utilize the computational resources available in cutting-edge GPUs. When memory bandwidth is the primary bottleneck rather than raw compute power, older GPUs remain surprisingly effective for inference tasks. The performance gap between newer and older GPU generations becomes much less significant than their computational performance might suggest.
Fig 2: GPU memory vs parameter count. Source: Gholani, Amir, et.al. (2024), AI and Memory Wall.
These technical characteristics create a unique hardware dynamic that changes the calculus around AI chips. Trailing-edge GPUs retain viability in an inference-dominated landscape — a generation-old GPU might deliver 60-70% of current-generation inference performance, making it highly viable for most applications. This shifts the cost-effectiveness equation; dollar-for-dollar, older GPUs often provide better inference performance per unit cost than cutting-edge hardware optimized for training workloads. While trailing-edge GPUs quickly become obsolete for training, they remain viable for inference much longer.
Architectural Innovations and Shifting GPU Viability
A single architectural innovation can reshape which GPUs are viable for inference tasks. DeepSeek's Multi-Head Latent Attention (MLA) highlights this dynamic, reducing KV cache requirements by over 90% and fundamentally changing inference bottlenecks.
By shrinking KV cache memory demands, MLA shifts short and medium-context inference tasks from being memory-bound to increasingly compute-bound. Lower memory demands mean GPUs spend less time waiting for data retrieval and more time on actual computation, significantly increasing GPU utilization rates. For China's AI ecosystem, this unlocks substantially more inference throughput from trailing-edge GPUs.
Custom optimizations further amplify these benefits. DeepSeek has demonstrated that Huawei's domestically-produced Ascend 910C can achieve 60% of Nvidia's H100 inference performance through targeted optimizations. This showcases how software and architectural innovations continually reshape the viability and relative strengths of different GPUs for AI workloads.
MLA renders short- and medium-context inference tasks far more efficient by reducing memory bottlenecks, allowing cutting-edge GPUs to fully leverage their computational power. While this widens the performance gap between cutting-edge and trailing-edge GPUs, it also increases China’s overall inference capacity by making older hardware more efficient. Leading-edge GPUs like the H100 will continue to dominate compute-bound workloads, but MLA significantly boosts the total inference power that can be extracted from China’s existing GPU stockpile.
For long-context inference, the hardware calculus shifts again. When context length becomes sufficiently large, tasks remain memory-bound even with MLA, reducing the performance advantage of cutting-edge hardware over trailing-edge hardware for these specific workloads. Long-context inference tasks are particularly important for reasoning, agentic AI, and automated research applications. The capacity of trailing-edge hardware to support these AI capability-enhancing tasks strengthens China’s ability to advance AI progress despite hardware constraints on cutting-edge GPUs..
Implications for Export Controls
The implications for export controls are significant: inference capacity is growing across the board, and restrictions on cutting-edge hardware won’t prevent China's inference capacity from expanding. Cutting-edge GPUs will retain significant performance advantages for short and medium-context workloads, but trailing-edge hardware remains surprisingly effective for long-context inference where memory constraints persist.
The prolonged viability of trailing-edge GPUs for inference extends the lifespan of China's existing hardware stockpile. Even as export controls limit China’s access to cutting-edge AI accelerators, China’s large stock of A100, A800, and H800 GPUs remains useful for inference applications far longer than they would for training. This sustains China's AI infrastructure and boosts its inference capacity despite limits on acquiring new chips.
Moreover, China has developed indigenous AI chips capable of inference. Huawei's Ascend 910C has demonstrated competitive performance for inference workloads. Notably, the Ascend 910C’s yield rate has doubled since last year to 40%, and Huawei plans to produce 100,000 units of the 910C and 300,000 units of the 910B in 2025, signaling a significant expansion of domestic chip production. Biren Technology's BR100, a 7nm, 77-billion transistor GPU, rivals the A100 for both training and inference. China’s growing production of inference-viable chips, substantial stockpile of trailing-edge GPUs, and continued access to the H20 reinforce its ability to sustain AI capabilities in an inference-heavy AI paradigm despite restrictions on acquiring cutting-edge hardware.
The Hardware Multiplier: China’s Inference Efficiency
Beyond hardware access, China’s advances in inference efficiency have significant strategic implications for U.S. export controls. DeepSeek’s recent innovations — particularly its v3 and R1 models — demonstrate China’s ability to push the frontier of inference efficiency. By implementing innovative techniques like a sparse Mixture of Experts architecture, multi-head latent attention, and mixed precision weights, DeepSeek’s R1 model achieves approximately 27x lower inference costs than OpenAI’s o1 while maintaining competitive performance.
This efficiency advantage effectively counterbalances U.S. hardware restrictions. Even if export controls limit China to 15x less hardware capacity, a 30x inference efficiency advantage would enable China to run nearly twice as much inference as the U.S. This acts as a multiplier on China’s hardware base, potentially giving China greater total inference capacity despite hardware restrictions.
The efficiency gains extend the utility of trailing-edge GPUs in China’s AI ecosystem, as improved inference efficiency compensates for computational and memory limitations. While DeepSeek’s achievements are a continuation of the observed decline in inference costs, this case demonstrates that Chinese AI labs have already developed the expertise to push the frontier of inference efficiency and could choose to withhold future breakthroughs if strategic considerations change.
The Sleeping Dragon: China’s Compute Overcapacity
Additionally, China’s massive but fragmented compute ecosystem is structurally better aligned with inference requirements than training needs. The aggressive GPU stockpiling during China’s “Hundred Model War” of 2023 created substantial compute capacity that became underutilized as many firms abandoned their foundation model ambitions. As Alibaba Cloud researcher An Lin observed, many of China’s claimed “10,000-GPU clusters” are actually collections of disconnected GPUs distributed across different locations or models. While this fragmentation makes the infrastructure suboptimal for training frontier models, it remains viable for inference workloads that can run effectively on smaller, distributed clusters.
Open-source models are particularly well-positioned to leverage this distributed infrastructure, enabling deployment across China’s fragmented GPU ecosystem and transforming previously idle compute into a strategic asset for widespread inference. This approach allows companies to preserve limited high-quality compute for model development while unlocking latent compute capacity.
China’s once-idle compute resources are increasingly valuable in an inference-heavy AI landscape, improving China’s position along both the innovation and diffusion axes.
How Should the U.S. Respond?
An inference-heavy AI paradigm favors China’s AI innovation potential. Its access to inference-viable hardware, leading inference efficiency, and compute overcapacity function better in an inference-driven context than in a pre-training one. U.S. export controls, designed to constrain training, have been less effective at limiting inference. China’s inference capacity remains underestimated. Despite restrictions, access to trailing-edge GPUs, stockpiles, domestic chips, and H20s enable continued progress.
As inference becomes central to AI competition, China’s relative position strengthens, narrowing the U.S. advantage. This shift demands a strategic recalibration: the U.S. must reinforce export controls and develop a coherent open-source AI strategy.
Restricting Exports of the NVIDIA H20
Export controls on AI hardware operate with a lag — typically one to two years before their full impact materializes. This lag effect is central to understanding both current policy outcomes and future strategic decisions for export controls.
Some cite DeepSeek’s latest models as proof that U.S. export controls have failed. However, this outcome is a shortcoming in how the controls were initially calibrated rather than a failure of the broader strategy. The Biden administration initially set narrow thresholds—based on FLOPs and interconnect bandwidth — which NVIDIA circumvented with the H800, designed specifically to remain exportable to China. When controls finally expanded to include the H800 in October 2023, Chinese companies had already stockpiled these GPUs in addition to speculated H100s and H20s, allowing them to maintain frontier development and delaying the policy's actual impact.
This lag highlights how AI hardware and model lifecycles can stretch over many months, so chips purchased immediately before or soon after a policy shift can remain in service for a long time. Consequently, the policy’s full impact may not be evident right away. As older hardware loses its edge for training and frontier development scales, the impact of controls becomes realized through constraints on both the speed of a country’s AI advancement and the extent of its diffusion.
The lag effect of export restrictions is more pronounced for inference hardware. Unlike training, inference workloads can remain viable on older GPU generations for much longer periods, as they depend more on memory capacity and bandwidth than raw compute power. If the U.S. delays restricting inference-oriented chips like the H20 until inference becomes even more central to AI power, the extended lag could substantially weaken the effectiveness of export controls as a defensive measure. By restricting the H20 now, the U.S. can meaningfully limit China’s accumulation of inference hardware before inference becomes the dominant compute paradigm in AI. The sooner these revised controls take effect, the sooner they will impose measurable constraints on China’s ability to compete along both axes of AI competition.
A Strategy for Open-Source AI
Open-source AI is a key vector of competition that requires a strategic U.S. approach. While it fuels innovation, not all models or circumstances warrant taking the same open approach. Open-sourcing an advanced model represents a form of technology transfer to China if that model exceeds the AI capabilities that China has access to. This reduces the U.S. lead on the AI innovation axis, shifting competition toward the diffusion axis — an area where China may be better positioned to compete.
As the compute requirements for pre-training grow, open releases help China overcome its pre-training disadvantage while amplifying the role of inference, where China is stronger. If not managed strategically, open-source AI could accelerate China’s ability to close the gap in both innovation and diffusion. The U.S. must assess whether it retains an edge in leveraging open models for research, application, and deployment. If so, open-source strategies can reinforce leadership; if not, they risk eroding it.
To assess the impact of an open release on U.S. tech competitiveness, we should evaluate how much of an immediate advantage the U.S. is foregoing on the AI innovation axis by open-sourcing a model and compare that to the net effect of how well the U.S. and China can convert open access into gains across both axes. If the U.S. retains a structural advantage in furthering AI research, building applications, fine-tuning, and scaling AI deployment, then open-source strategies can reinforce U.S. leadership. However, if China is more effective at leveraging open models for research, real-world adoption and economic or military applications, then unrestricted open release could benefit China more. This dynamic underscores the need for a structured approach and collaboration between private and public sector regarding deployment decisions.
The Bottom Line
As trends in AI elevate the importance of inference, the U.S. must reassess its strategy to lead along both axes of AI competition. While early export controls are designed to constrain China’s ability to train frontier models, they are less effective in limiting its capacity for large-scale inference. To sustain its competitive edge, the U.S. must expand export controls to address the growing role of inference, particularly by restricting chips like the NVIDIA H20 before their strategic importance escalates further. At the same time, the U.S. must refine its approach to open-source AI, ensuring that its diffusion benefits reinforce, rather than undermine, U.S. national AI leadership. Winning the AI competition requires adapting as fast as the technology evolves, and this is a critical moment for the U.S. to recalibrate its strategy.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
It’s been a tough week for the international order. It feels like every TV in every restaurant across Taiwan is blasting nonstop coverage of the Trump-Zelenskyy fallout.
How will Taiwan respond to Trump’s pivot to Putin? Would Taiwan be safer with nuclear weapons? What platforms do Taiwanese people use to debate about politics anyway?
In today’s roundup, we’ll analyze perspectives from Taiwanese legacy newspapers, social media firestorms, and viral political influencers.
Driving Solidarity
We’ll start off by highlighting some reactions on the most popular Taiwanese social media platform, PTT.
PTT is a bit like a Taiwanese version of Reddit. The key difference is that comments are always displayed in chronological order instead of being ranked by popularity. Users can “push” 推, “boo” 噓, or reply to comments to express their opinion. The platform shows whether each comment is being “pushed” or “booed” overall, but doesn’t display the total vote tallies. Like on Reddit, there are sub-forums for topic-specific discussion.
Disclaimer: these forums are hosting open debates with intense back-and-forth between commenters. I’ll be highlighting recurring themes, as well as arguments where both sides are earning push-votes, but I want to be clear that there is no broad consensus on what the Trump-Zelenskyy fallout means for Taiwan at this point.
(Pushed) I support and praise this article, justice will prevail.
(Booed) Then how come there are no soldiers? Conduct an opinion poll or something.
(Pushed) These past couple of days, I've seen quite a few people claim that Ukraine should have originally surrendered to Russia in exchange for peace and prosperity. This kind of argument completely ignores the suffering Ukraine endured under Russian rule in the past.
(Pushed) In the past, we thought that people in democratic countries feared death more than other people — but Ukrainians are not afraid.
(Pushed) The Uyghurs will never surrender, but they will not go to the front line
(Pushed) It is 100000000% reasonable to be suspicious that Trump received personal benefits from Russia or made a blood pact with Russia.
The Taiwanese transliteration of “Zelenskyy” is 澤倫斯基 Zélúnsījī, and in casual writing Taiwanese people refer to him by the nickname 司機 Sījī (literally, “The Driver”) which has the same pronunciation as the last two characters of the transliteration.
From a thread in a military forum about whether Zelenskyy overplayed his hand:
(Pushed) The driver really shouldn’t have talked back to Vance. If he wanted to argue, he could have done it in private.
(Pushed) After apologizing, you still have nothing, so why bother apologizing?
(Pushed) If Little Z doesn’t kneel, America will make explosive corruption accusations against him.
(Reply) East Asian countries are better at licking.
(Pushed) If Ukraine wants to thank someone, it should thank the previous Biden administration. Why thank Trump?
(Pushed) It seems someone is trying to smear and destroy Mr. Z's image. Be careful when responding to this thread.
Indeed, there are signs of disinformation in some discussions of this topic. An FT article entitled “Zelenskyy rejects calls for immediate Ukraine-Russia ceasefire” was posted on PTT with the mistranslated title, “The Driver Rejects Ukrainian and Russian calls for a Ceasefire” (司機拒絕烏克蘭與俄羅斯立即停火的要求), a fact which was quickly pointed out and mocked in the comments.
Marco Rubio is well-known in Taiwan thanks to his long congressional record of support for the island. Here are some comments about him:
(Pushed) Rubio will be replaced soon.
(Pushed) Rubio was once a pioneer in anti-communism, but now he bows down to power.
Underneath an article reporting Trump’s plan to freeze aid to Ukraine in response to the meeting:
(Pushed) Stop it right now immediately!!!!!!!!!!!!!!! I’ve never seen such a cowardly U.S. president!! You truly see everything if you live long enough!!!!!!!!
(Pushed) Will the European big brothers shoulder some of the responsibility? Isn’t this an opportunity for them to show off?
(Pushed) Being pro-China is selling out Taiwan, being pro-America is also selling out Taiwan.
(Pushed) In the Budapest Agreement, even China said it would protect Ukraine, but that isn’t happening
(Pushed) Ultimately, [Ukraine] should not have given up its nuclear weapons. Security guarantees are bullshit.
(Booed) Ukraine has no nuclear bombs, so of course it has no bargaining chips.
(Pushed) The driver’s bargaining chip is making the king (Trump) lose face.
(Pushed) Buddha’s mercy 佛祖慈悲 [This phrase is used ironically in situations that are cruel or corrupt to the point of hopelessness.]
Ukraine Today, But Taiwan’s OK?
At the start of the invasion, the DPP popularized the slogan, “Ukraine today, Taiwan tomorrow.” Editor Gu Shu-ren 辜樹仁 of CommonWealth Magazine 天下雜誌 (a Taiwanese publication similar to the Atlantic), addressed fears that Trump will abandon Taiwan after Ukraine in a recent editorial:
Looking back at history, Taiwan's strategic value to the United States has been the key factor in America's decision to either abandon or support Taiwan.
In 1950, when the Korean War broke out, the Republic of China (ROC) government, which had retreated to Taiwan and was on the brink of collapse after being abandoned by the U.S., suddenly became the central hub of the U.S. first island chain strategy in East Asia — a so-called unsinkable aircraft carrier — greatly increasing Taiwan's strategic importance.
In the 1970s, as the U.S. aligned with China to counter the Soviet Union, Taiwan lost its strategic value, leading to the severance of U.S.-Taiwan diplomatic ties and the withdrawal of U.S. troops from Taiwan. …
Today, Taiwan's strategic value to the United States is at its highest since the servering of diplomatic ties, as the primary battleground in the U.S.-China rivalry is now the technology war, with semiconductors at its core. More specifically, TSMC is the most crucial asset for the U.S. in securing a supply of advanced chips and revitalizing its semiconductor manufacturing industry. If the U.S. wants to maintain its technological and military lead over China, it must firmly keep Taiwan within its grasp. …
Ensuring that the U.S. remains dependent on Taiwan’s advanced chip manufacturing — making American national security synonymous with protecting Taiwan — is the most critical factor in maintaining Taiwan’s strategic value to the United States.
Of course, there is another equally important factor. Trump dislikes war, especially costly military interventions where the U.S. cannot be assured of victory. He has repeatedly complained that Ukraine failed to prevent war at the outset. Therefore, avoiding war at all costs is also a key strategy for Taiwan to secure Trump’s support.
Only through this can tomorrow’s Taiwan avoid becoming the Ukraine we saw today.
Reporter Jiang Liangcheng 江良誠 similarly warned that Taiwan would need to become more transactional in its relationship Trump:
“Trump's only vocabulary is actually "money, money, money". All international relations can be measured by money. There is no free lunch in the world. It is impossible to ask Americans to help you defend your country like a plate for free and without any reward. …
However, when it comes to Taiwan's policy toward the United States, Lai Ching-te still sticks to Tsai Ing-wen's international politics, such as the first island chain, geopolitics, and Indo-Pacific security. I'm afraid even Trump doesn't understand these terms.”
The Meihua News Network (梅花新聞網), a Pro-China news outlet owned by a controversial Taiwanese religious leader, argued instead that Taiwan needs to reopen dialogue with Beijing given the reality that the U.S. is an unreliable partner.”
In front of cabinet members and the media, Trump was unwilling to guarantee that the Chinese Communist Party would not invade Taiwan by force during his term, and emphasized that he had a good relationship with Chinese Communist Party leader Xi Jinping. …
“Foreign Affairs” recently published a special article titled “The Taiwan Fixation: American Strategy Shouldn’t Hinge on an Unwinnable War”, co-authored by Professor Kavanagh of the Georgetown University Center for Security Studies and senior scholar Wertheim of the Carnegie Endowment for International Peace. The gist of the article is: Taiwan is certainly valuable to the United States, but if American decision-makers overestimate Taiwan's importance, they will sacrifice the security of maintaining the status quo due to the risk of endless and destructive war; and Taiwan's importance is not enough for the United States to sacrifice tens of thousands of American lives to protect it. Former National Security Council Secretary-General Su Chi 蘇起 described this article as the most powerful article to date advocating the United States to let go of Taiwan. …
Apart from fully relying on the American security umbrella and turning Taiwan into a "porcupine," the DPP also has another option: restoring cross-strait communication and reducing tensions in the Taiwan Strait. If that happens, the so-called "Abandon Taiwan Theory" would naturally dissipate. Rational decision-making should not be obstructed by anti-China or China-hating sentiments.”
By contrast, a popular post from the Taiwanese political influencer James Hsieh argued that Taiwan should be doing whatever it takes to improve relations with the U.S., not criticizing Trump’s Ukraine policy:
“I still see many people online going against the tide, bashing Trump, criticizing the U.S., and supporting all kinds of conspiracy theories. Here are five reminders:
Before the war, Ukraine was extremely pro-China, selling major military technology to China. Just a few days ago, Ukraine even asked China for help.
Morally, we must oppose aggression, but in terms of international strategy, we must firmly support the United States.
Taiwan is not Ukraine. In terms of historical ties with the U.S., the Taiwan Relations Act, geographical location, type of warfare, and economic strength, Taiwan is completely different. Taiwan is absolutely not a distant European country like Ukraine in America's eyes. Comparing Ukraine to Taiwan is a completely flawed analogy. Saying that the U.S. pulling out of the Russia-Ukraine war implies that it will betray Taiwan is just another favorite conspiracy theory of the dumb lefties (左膠) and the Chinese Communist Party’s propaganda machine.
Personally, I hope the Russia-Ukraine war ends quickly so that the U.S. can fully prepare for the Indo-Pacific. This is a practical concern, as China is rapidly advancing its strategic plans. How the U.S. swiftly ends its engagements elsewhere and refocuses on the Indo-Pacific is critical. Just yesterday, Vice President Vance stated that the U.S. military-industrial production can no longer sustain the continuous supply of heavy weaponry to Ukraine.
History has shown that during major wars, opportunistic nations take advantage of a great power’s exhaustion to invade smaller neighboring countries. If the Russia-Ukraine war escalates into World War III and the U.S. and Europe are preoccupied with fighting Putin’s alliance, it would be the perfect moment for China to seize Taiwan under the guise of maintaining stability.
If Taiwan's democracy, freedom, and independence from oppression are what you value most, then Taiwan should prioritize its relationships with the U.S. and Japan over everything else — not Ukraine.
Only the U.S. and Japan will help us. Survival comes first before ideals.
Taiwan-U.S. friendship!”
It remains unclear what the Lai administration’s approach will be, but you can be sure that ChinaTalk will keep monitoring the debate as it evolves.
Zelenskyy’s White House press conference also reignited the olddebate about whether Taiwan would benefit from having its own nuclear arsenal. Taiwan abandoned its indigenous nuclear program in response to pressure from the U.S., much like how Ukraine relinquished its nuclear weapons to Russia after the fall of the USSR. Taiwan was estimated to be just two years away from completing a WMD when the U.S. intervened in 1988.
These parallels were drawn explicitly by a CNN profile of Colonel Chang Hsien-yi 張憲義, the Taiwanese nuclear engineer who provided intelligence about Taiwan’s proliferation plans to the CIA. The article was repackaged, translated, and published on the front page of the China Times on Monday.
(Pushed) This person is the reason why Taiwanese independence is impossible.
(Pushed) Nuclear weapons are not something that Taiwan's extremely incompetent politics could handle. If nuclear weapons were in the hands of Chiang Kai-Shek and his family, Taiwan would have ended up like North Korea. The Chiang family would still in power, and there would never have even been a chance for democratization. So many people have no clue what’s going on.
Taiwanese political influencer Mr. Shen 公子沈, who runs a YouTube channel with more than 700k subscribers, posted the following meme on Threads (which is way more popular in Taiwan than the U.S.) with the caption, “With nukes vs without nukes: it’s time for Taiwan to develop nuclear weapons.”
Speaking of bargaining chips…
Reactions to the TSMC Deal
TSMC’s newly announced $100 billion investment in US chip manufacturing led to more online discontent. The following comments from Facebook were curated by Angela Oung:
“So they’re taking our stuff, leaving us with no cards. Think they’ll help in the future? Stop dreaming!”
“Taiwan’s remaining value is becoming a meat grinder like Ukraine.”
“He [TSMC Chairman CC Wei] looks like he has a gun behind his head. Hostage situation.”
“The silicon shield we spent decades building is being handed over by our government without a whimper”
“TSMC: built by the KMT, sold by the DPP”
“Is Lai Ching-te such a pussy that he’s not even gonna say anything?”
“Today Ukraine, tomorrow Taiwan. One step closer to refugee status.”
“Bandits…just like the CCP”
To close, I’ll leave you with another popular post on Threads expressing frustration about Taiwan’s-U.S. relations:
“The U.S. asks us to buy military equipment — we buy it.
The U.S. asks us to extend the length of mandatory military service — we extend it.
The U.S. wants TSMC — we hand it over with both hands.
The U.S. wants us to implement resilient defense — we manage to do it, even if we have to hide and shuffle the budget.
For every single thing the U.S. asks of us, from the issue of eating ractopamine pork in our daily meals to national defense policies involving regional security cooperation, Taiwan follows the U.S.’s demands without question.
But will there come a day, just like today’s Ukraine, where we sign agreements on resource concessions, trading away our country's future rebuilding assets, yet still lack the most basic “security guarantees”?
Ukraine has the support of the entire European continent—but what about Taiwan?
Will today’s Ukraine be a reflection of Taiwan’s future?
Will Taiwan, when that day comes, be even more isolated and helpless?”
To be fair, this commenter is right that Taiwanese pork is way more delicious than the ractopamine pork imported from the U.S. I sincerely hope that every ChinaTalk subscriber has an opportunity to come to Taiwan and eat stewed pork rice (滷肉飯)…before it’s too late!?
Source. Jordan does not eat pork and does not approve this message.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
你在学习哪些能让人生更美妙的东西呢？它吸引你的原因是什么呢？学习的方法和最主要的收获（Learning and key takeaway）是什么？最让人惊讶新知又是什么呢？欢迎来稿你的分享——建立一个自愿的终身学习乐园，每一个学习者也都是知识的传播者和分享者。投稿发送afterschool2021@126.com。
你在学习哪些能让人生更美妙的东西呢？它吸引你的原因是什么呢？学习的方法和最主要的收获（Learning and key takeaway）是什么？最让人惊讶新知又是什么呢？欢迎来稿你的分享——建立一个自愿的终身学习乐园，每一个学习者也都是知识的传播者和分享者。投稿发送afterschool2021@126.com。
A guest piece by Afra, freelance writer and podcaster [Jordan: I highly recommend this show!] with working experience in tech and crypto. Personal site here.
DeepSeek’s winds have already been blowing for some time, but this particular gale seems to have real staying power.
On Chinese social media, the discussions took on a life of their own, with the most popular use case being the calculation of one’s Ba Zi (八字) and astrological chart, using the social media tag “AI玄学” (AI Mysticism). Users weren’t just seeking their personal fortunes — they saw the nation’s destiny itself shifting through DeepSeek’s emergence. These conversations are a swirling mix of collective jubilation, national pride, and gleeful satisfaction over America’s “China envy,”1 often accompanied by playful banter.
Yet amidst this discourse, a deeper and more resonant question emerges: could this be a sign of China’s technological ascension? Is this evidence that Guoyun (国运) — the nation’s long-awaited destiny — has finally arrived?
First, what is Guoyun 国运？
The term 国运 combines two characters: 国 (guó, “nation/state”) and 运 (yùn, “fate/destiny/fortune”). This concept emerged from traditional Chinese cosmological thinking, where the destiny of the state was seen as intertwined with celestial patterns and dynastic cycles.2 This term, once confined to the ornate dialogue of period dramas set in imperial China, has begun to surface with increasing frequency on my social media timeline.
For Chinese netizens, discussions about politics on social media are often marked by subtlety and veneration with trepidation (for reasons that require little explanation). However, during the 2025 Chinese New Year, the discourse expanded far beyond politics and DeepSeek into a cacophony of cultural euphoria —a wave of self-congratulatory enthusiasm that evolved into something larger culturally. This included the movie Nezha 2, which shattered box office records and surpassed Inside Out 2 to become the highest-grossing animated film of all time (with patriotism-fueled consumption boosting the box office performance), TikTok refugees flooding Xiaohongshu, and advanced Unitree robotics performing during the Spring Festival Gala. These achievements seemed to occur against a historical backdrop where technological and cultural advances carry deeper significance about China’s rightful place in the cosmic order.3
Screenshot of a typical post on national destiny. The first comment says: “I hope my luck can take off like the national destiny.” The second comment says: “Why is everyone so shocked [about DS]? China is not the number one in the world for only 1-2 hundred years, and China has worked so hard during this period. Isn’t normal for China to achieve its goal?”
The Guoyun discourse extends beyond tech leaders, media commentary, and social media posts.
President Xi Jinping has woven the concept of destiny into official rhetoric, though carefully stripped of its more superstitious elements. Speaking at the 19th Academician Conference of the Chinese Academy of Sciences in May 2018, Xi declared, “Innovation determines the future; reform concerns national destiny. The field of science and technology is the area most in need of continuous reform 创新决胜未来，改革关乎国运。科技领域是最需要不断改革的领域.” This statement aligns with his broader techno-nationalist vision, explicitly linking technological advancement to China’s strategic future.
A 2024 People’s Daily article discussing Xi’s thoughts emphasized that “cultural confidence is a major issue concerning national destiny 坚定文化自信，是一个事关国运兴衰...的大问题"。
This rhetorical shift signals a carefully calibrated blend of traditional Chinese concepts with modern governance — a bridge between ancient ideas of dynastic cycles and contemporary aspirations for technological supremacy.
Beyond superstition: is this a collective myth-making or post-pandemic yearning for certainty?
It would be a mistake to dismiss this discourse as mere superstition or propaganda.
The COVID-19 pandemic marked a watershed moment in Chinese society’s relationship with national destiny. To me, Zero COVID became a mirror polished to cruel clarity, reflecting a China I no longer recognized. During the rigid cycles of lockdowns and reopenings, I didn’t see my parents for two years, my grandmother was hospitalized, and my cousin was confined to his university dorm for three whole months culminating in a severe mental breakdown. Friends lost loved ones due to a lack of timely treatment options. Back then, seeing how waves of people wanted to “run (润)” from China, I thought for the first time that I might never return to China, and that I might become part of the Chinese diaspora forever.
COVID created a collective trauma that many Chinese are still processing.
But this experience has paradoxically reinforced a certain earnest faith in China’s future among ordinary citizens. The optimism in the discussion of Guoyun might represent a complex emotional response to the uncertainty and trauma from the COVID era — a blend of traditional fatalism with genuine aspirations. Having weathered the pandemic’s disruption, many ordinary Chinese seek reassurance about the future through familiar cultural frameworks. ‘National Destiny’ provides exactly that — it’s a narrative that contextualizes current struggles within a larger, ultimately triumphant story. It’s therapeutic.
The discourse around 国运论 (guóyùn lùn, or “national destiny theory”) reveals parallels to America’s historical myth-making. Perhaps the most striking similarity between China and the US is their unwavering belief in their own exceptionalism and their destined special place in the world order. While America has Manifest Destiny and the Frontier Thesis, China’s “national rejuvenation” serves as its own foundational myth from which people can derive self-confidence. Through countless repetitions across state and social media, this narrative has become deeply ingrained in China’s national consciousness.
The wounds behind techno-nationalism
Where myths nurture the national consciousness, technology has become the battleground where China’s historical narrative demands its vindication. The roots of China’s techno-nationalism run deep, drawing emotional power from China’s “century of humiliation.” U.S. actions — chip controls, the attempted TikTok ban, tariffs, investigations of Chinese scientists, and suspicions of Chinese espionage — rekindle the historical trauma of humiliation.
For decades, China has been portrayed as a mere copycat or thief of Western innovation. Each technological breakthrough now serves as vindication, a refutation of that dismissive narrative — this shame has never truly been resolved. As Kevin Xu elaborated on DeepSeek’s open-sourced nature, “It’s all for the validation and approval,” — a sharp acknowledgment that when Chinese engineers share their code with the world, they’re not just demonstrating technical prowess but seeking to heal a wound in the national psyche:
In the Chinese open source community, there is this thing that I would call open source “zeal” or “calling” (开源情怀)
Most engineers are thrilled if their open source projects — a database, a container registry, etc-- are used by a foreign company, especially a silicon valley one. They’d tack on free labor on top of already free software, to fix bugs, resolve issues, all day all night. It’s all for the validation and approval.
Implicit in this “zeal” or “calling” is an acute awareness that no one in the West respects what they do because everything in China is stolen or created by cheating. They are also aware that Chinese firms have been taking for free lots of open source tech to advance, but they want to create their own, contribute, and prove that their tech is good enough to be taken for free by foreign firms -- some nationalism, some engineering pride.
So if you want to really understand why DeepSeek does what it does and open source everything, start there. It’s not a political statement, not to troll Stargate or Trump inauguration, or to help their quant fund’s shorts on NVDA (though if that were the case, it’d be quite brilliant and savage)
The drive to prove oneself on behalf of the nation is expressed vividly in Chinese popular culture. I couldn’t stop thinking about Illumine Linga (临高启明), an open-source collaborative novel that has captivated China’s engineering community and become a phenomenon of its own. The story follows modern Chinese engineers who time-travel to the declining Ming dynasty, right before China was conquered by the Manchus, bringing industrial equipment and technical knowledge. They gradually industrialize Hainan and Guangdong provinces before expanding outward with the ultimate goal of establishing global hegemony.4
A screenshot of an online forum dedicated to Illumine Linga. The front page features DeepSeek’s founder, Liang Wenfeng, as he resembles a character in the novel.
Though ostensibly just fiction, Illumine Linga pulses with the heartbeat of China’s “Industrial Party” (工业党) — that loose constellation of engineers, programmers, and technically-minded patriots united by an almost religious faith in technology as destiny’s instrument. The novel serves as a sharp allegory for contemporary aspirations: technological mastery as the path to national resurrection and global respect.5
In the Western intellectual tradition, technology and data have undergone phases of detached scrutiny — viewed first as tools of emancipation, and later as vectors of control. Foucault’s panopticon mutated into Zuboff’s surveillance capitalism; Wiener’s Cybernetics birthed both Silicon Valley and Snowden’s disclosures. This academic back-and-forth assumes a fundamental premise: technology can theoretically exist as a neutral substrate awaiting ideological imprint.
However, in my impression, China’s techno-discourse never evinces such “purity.”
From its inception, technology has been semantically encased in the shell of techno-nationalism. In China’s history textbooks, Qian Xuesen’s missiles for the Two Bombs, One Satellite program were never just missiles, but brushstrokes in the narrative of “standing up again.”6 Yuan Longping’s hybrid rice strains didn’t merely feed millions; they were genetic correctives to the “Century of Humiliation,” each harvest a quiet refutation of the colonial-era belief that China couldn’t innovate.
On Chinese New Year’s Eve, a fake response to the “national destiny theory” attributed to Liang Wenfeng circulated widely online, with many believing and sharing it as authentic. This response claimed that DeepSeek’s open-source decision was merely “standing on the shoulders of giants, adding a few more screws to the edifice of China’s large language models,” and that the true national destiny resided in “a group of stubborn fools using code as bricks and algorithms as steel, building bridges to the future.” This fake statement—notably devoid of wolf warrior rhetoric—spread virally, its humility and relentless spirit embodying some values people hoped Chinese technologists would champion. Meanwhile, the real Liang Wenfeng remained silent after DeepSeek’s rise. A month later, he appeared on CCTV sitting beside Tencent’s Ma Huateng at Xi Jinping’s symposium for top business leaders.
The public’s fascination with Liang showed no signs of waning. In Silicon Valley, his previous interviews were swiftly translated into English and meticulously analyzed, while in China, his rise also inspired mystical interpretations—during the Spring Festival holiday, Liang Wenfeng’s ancestral home in Zhanjiang, Guangdong transformed into an impromptu tourist attraction, drawing feng shui masters eager to study the geomantic properties of his family residence.
Humans have always sought ways to calculate the incalculable. Perhaps that’s what makes the conversation around Guoyun so captivating: it’s not just about predicting the future, but about sense-making in China’s present.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
I will skip other related concepts about “national destiny,” including how Chinese emperors employed court astrologers, consulted the I Ching, and the concept of the Mandate of Heaven.
Additional signs of China’s 国运 emerging include the new marriage law (which broadly supports women’s rights and economic independence), the global success of “Black Myth: Wukong,” NeZha 2’sa performance at the box office, and the Spring Festival Gala featuring more diverse and open programming than in previous years, indicating some deeper vibe shift.
As Illumine Linga has grown in length, this collaboratively written novel has expanded to encompass diverse themes: women’s rights, Marxism, power struggles, military strategy, and aesthetics, among many others…And of course, public reception to the novel is diverse. Some Chinese readers find it embarrassingly nationalistic, while others dismiss its premise as simplistic fantasy. It’s worth noting that this work doesn’t represent universal sentiment—large segments of China’s tech community remain either unaware of Illumine Linga or view it with skepticism rather than admiration. But again it does captures the validation-seeking mentality so precisely.
Tianyu Fang wrote a piece showing how Qian Xuesen’s departure from the U.S. and service in China was inevitably geopolitical. Qian’s “return” also became part of an official nationalistic narrative that has persisted for decades.
Gary Wang spent the past decade developing business and product strategy for Silicon Valley technology companies, with a focus on enterprise software, the industrial internet of things and AI. He has a degree from HKS and worked in China. The views expressed here represent only his own.
About a decade ago, the best forecasts heralded a promising manufacturing future, in the United States and globally, with the advent of the fourth industrial revolution (also called “industry 4.0,” the “industrial internet,” or “industrial internet of things” aka IIoT). The belief was that the falling cost of cloud computing, sensor costs, and machine learning — coupled with new connectivity technologies such as 5G or IPv6 — would lead to a revolution in manufacturing productivity and ultimately higher GDP growth.
Despite these promising forecasts, multiple data points indicate that US manufacturing has largely stagnated. Analysis from the New York Federal Reserve reveals that both total factor productivity and labor productivity have been flat from 2007 to 2022. Meanwhile, US share of global manufacturing value add fell from nearly 25% in 2000 to an estimated 15% today in 2024. The UN Industrial Development Org projects US share of global manufacturing value add will fall to 11% in 2030, while China may account for 45% of global output.
This decline comes after multiple presidential administrations’ efforts to revitalize American manufacturing — from the Obama-era policies such as the Advanced Manufacturing Partnership or the Manufacturing USA initiative, to the Biden administration’s Inflation Reduction Act, and now the Trump administration’s desire to reshore manufacturing via tariffs and other policy tools.
Off-shoring and free-trade agreements go only so far in explaining this decline. And the present debates over US industrial policy — sparked by the advent of emerging technologies (generative AI, quantum computing) as well as intensifying competition with China — perhaps focus on the wrong things.
The real questions US policymakers must grapple with: why did the United States fail to capitalize on technology that was already available to make its manufacturing base more competitive?
Put another way: why have the promises of the IIoT revolution failed to materialize in the United States?
This piece makes a few key arguments:
The “industrial internet of things” is not an industry. It’s a set of disparate technologies that all need to be adopted together to create value.
The free market will not always optimize adopting a broad set of technologies for an entire ecosystem of industries. The underwhelming results of today’s industrial internet is a case in point.
China’s industrial policies to “win” the fourth industrial revolution offer lessons for policymakers in the United States to consider.
When it comes to revitalizing manufacturing, or ensuring American leadership in AI or quantum computing, policymakers need to craft policies to develop entire value chains and tech ecosystems — not myopically focus on just one strategic technology (eg. advanced semiconductors).
ChinaTalk is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
What is IIoT?
IIoT refers to the interconnection of machines, devices, sensors, and systems which are connected on the internet in performing industrial tasks.
Take one of IIoT’s leading “use cases” (ie. applying tech to solve a business problem): predictive maintenance. Sensors connected to a piece of factory equipment, such as a boiler, can measure temperature or vibration. When combined with machine-learning algorithms, manufacturers stand to save millions by predicting when a machine would fail, and then proactively maintaining the machine before a failure occurs — thus reducing factory downtime and increasing productivity.
Another use case: gathering GPS data from truckers could enable machine-learning algorithms to optimize the routes of commercial trucks (saving fuel costs). When paired with data on customer demand (say, Pepsi sales in a city), manufacturers could save billions by optimizing their inventory costs to ensure that the optimal amount of Pepsi reached store shelves at just the right time.
The use cases are endless: deploying robots on the production line, using cameras and AI to automate quality inspection for finished products, creating a “digital twin” of an entire production process for optimization, and much more. All of these use cases required cloud computing, real-world historical data, and connectivity. As a practitioner who has worked with technology companies on their strategy for delivering industrial IoT to manufacturing companies, I can attest to the level of industry enthusiasm for IIoT during this time (as well as the numerous operational challenges).
How off were the IIoT forecasts?
In 2015, McKinsey forecast $1.2 to 3.7 trillion in economic value created per year by 2025 from IoT technologies in factories. Assuming technology vendors alone capture 5% of the value created — a very conservative benchmark — that’s $60 to $185 billion in revenue. The International Data Corp in 2017 forecast that manufacturers would spend $102 billion in the industrial internet, meaning vendors selling IIoT technologies should see comparable revenue figures. Accenture and World Economic Forum joined the hype, intoning that the “Industrial Internet will transform many industries, including manufacturing, oil and gas, agriculture, mining, transportation and healthcare. Collectively, these account for nearly two-thirds of the world economy.” These market forecasts led the Congressional Research Service in 2015 to predict, “The current global IoT market has been valued at about $2 trillion, with estimates of its predicted value over the next 5 to 10 years varying from $4 trillion to $11 trillion.”
These forecasts were off by multiple orders of magnitude. Today, to my knowledge, there is only one publicly listed company in the United States solely focused on IIoT: Samsara, with $1.4 billion in revenue, growing at a healthy ~40% year over year. (Palantir in 2024 reported $700 million in revenue from US private-sector firms, some of which include manufacturing — but the majority of Palantir’s business is with governments.)
General Electric and Siemens both tried to become technology companies by developing their own cloud platforms and AI applications to digitize the manufacturing sector. A series of New York Times headlines, though, tells the saga of GE’s attempt to capture the purported massive opportunity of the industrial internet of things:
Siemens positioned its industrial internet cloud platform, Mindsphere, as its next growth vector. Today, Mindsphere has been rebranded to “IoT insights hub,” and the last time Siemens company leadership talked about Mindsphere on their earnings call with equity analysts was in 2022, indicating a retrenchment in expectations (unlike when they spoke about Mindsphere on earnings calls with analysts in 2015, 2016, 2019, and 2020; what industry leaders tell Wall Street indicates where they think their companies’ growth will come from).
Why were the predictions so wrong?
IIoT is a cluster of disparate technologies that have to work together to create value. It’s not one technology. Consider the aforementioned predictive maintenance use case. To realize value, a factory owner needs to adopt six or seven different technologies from different vendors.
There’s the company providing sensors (sometimes with software) for the machines to gather data for analytics.
Many factories have historically not been connected to the internet, so a company like Verizon needs to get involved to set up an in-plant 5G connectivity network. (Leading analysts have estimated there are only a handful of 5G industrial projects in the United States, compared to likely thousands in China.)
A company like Cisco has to provide the networking equipment to enable internet connectivity in the factory.
A cybersecurity company needs to ensure the sensors and machines, now that they’re connected to the internet, are secure from cyberattacks.
A cloud-computing company, such as Microsoft or Amazon, needs to provide the compute and storage for the customer to develop AI algorithms to analyze the data generated by the sensors. These cloud-computing companies often provide the AI algorithms for customers to customize themselves (assuming they have the in-house data science talent) to analyze the data from factory equipment.
A company needs to integrate these disparate systems together — usually a system integrator like Accenture or Wipro.
The factory owner has a finite budget, must negotiate with six different vendors (each with their own pricing and profit models, none of whom necessarily coordinate their selling activities) — but still must realize a high enough return on investment (ROI) to justify solving this one use case. Imagine a consumer buying a car — but instead of buying from an OEM like Tesla or General Motors, you have to negotiate individually with the tire company, the engine manufacturer, the seat belt maker, the company making the infotainment display, and every other component manufacturer.
The nature of the physical world makes this coordination problem even more complex:
Algorithms aren’t immune from false positives. What happens if the algorithms incorrectly predict a machine will break down, but a maintenance technician has already been dispatched to make repairs? That reduces ROI.
Machine algorithms need to be trained on historical data of when the machine has broken down before — but for many factories, maintenance records aren’t digitized; if available at all, they’re paper logs of when a technician fixed a machine.
Third, from the perspective of the technology vendor, sales cycles to manufacturers often are usually one to two years — since customers will pilot the technology for one set of machines (one use case) in one factory, measure the cost or productivity savings, and then decide whether they want to scale the technologies to multiple use cases across multiple factories. Factory budgets are managed locally, not globally — meaning a vendor has to sell to a manufacturer’s factory site in, say, the United States, then Brazil, then Germany, and so on.
All of these factors help to explain why venture capitalists — with few exceptions — have not invested in startups tackling industrial IoT, as well as why it’s been hard for existing vendors to scale their business. Even McKinsey admitted in 2021, “To date, value capture across settings has generally been on the low end of the ranges of our estimates from 2015, resulting from slower adoption and impact. For example, in factories, we attribute the slower growth to delayed technological adoption because many companies are stuck in the pilot phase.”
What has China done?
While IIoT hasn’t lived up to its potential in the United States and elsewhere in the West, China has leaped ahead in the fourth industrial revolution: there is no other country in the world that can boast of legions of “dark factories” — ie. factories where entire manufacturing processes are automated.
How has China done it? By focusing on technical challenges and market-coordination problems.
First: Chinese policymakers at the highest level — eg. the State Council — crafted policies to solve known technical challenges which threatened to hold back Chinese manufacturer’s adoption of IIoT technologies.
For example, in the predictive maintenance use case, there is a known problem of “asset mapping” — ensuring all the physical and digital assets in a factory can be identified in a common taxonomy to enable machine-learning analytics and then workflow automation (sending a technician to repair a robot, changing the workload of robots working together if one robot is breaking down, etc.). Specifically, if factory owners want to predict when a robot arm will break down, they need a comprehensive way to uniquely identify the specific robot, the specific arm of that robot, the specific sensor that may be attached to the robot, the specific 3D model of the robot’s arm, and then map all of these physical and digital assets together. Without a common taxonomy, it’s impossible to automate the analysis of sensor readings from the robot arm (eg. its grip strength) and then trigger a workflow to fix the robot arm while enabling the production process to continue seamlessly, that is, in a “lights out” fashion.
China’s State Council, in a 2017 planning document — “Guidance for Deepening the Development of the Industrial Internet ‘internet + promoting manufacturing” 深化“互联网+先进制造业” 发展工业互联网的指导意见 — specifically called for implementing networking connectivity and “identity resolution system” 标识解析体系 to solve this problem, using a combination of known technologies and standards such as IPv6, software-defined networking, 5G connectivity, time-sensitive networking, and passive optical networking. The technologies mentioned in this document were available in China (and the United States) in 2017. An identity resolution system (the English equivalent term would be a digital “tracking system”), when combined with advanced networking technologies, solves this predictive-maintenance problem because then a piece of software — such as a predictive-maintenance application for robots — can automatically locate the robot arm that’s emitting sensor data indicating a breakdown, match that to the 3D model that specifies how the robot arm should function, detect issues with the robot arm, and then trigger a workflow to remediate. Dozens of physical and digital systems are involved in solving this problem.
Of course, the free market can solve this problem as well — but it runs into the same issue mentioned above: coordination of multiple vendors with multiple technologies and standards that all have to work together. No wonder that, in 2024, 5G adoption in the US manufacturing sector was at 2%. After all, a factory doesn’t realize any business value from just deploying 5G by itself, if the rest of the technology stack (sensors, algorithms, applications, cloud computing, security, etc.) isn’t also deployed.
Second: China targeted industrial policy to solve known market-coordination problems that would hold back IIoT adoption.
For example, consider the problem of sub-scale platforms. To better understand what this is, I’ll first lay some foundation on key terms:
A platform is any technology in which an underlying resource, such as computing power (eg. Amazon Web Services), is offered to customers as a software component to build a fully functional piece of software. In the IIoT case, “industrial internet of things platforms” are cloud platforms that allow manufacturers to (1) access compute and data storage, (2) enable data to be sent from physical machines to the cloud, and (3) secure the network and data from machine to cloud. An IIoT application is a packaged piece of software with algorithms and an end-user interface that solves a business problem.
The consumer analogy is how the iPhone is a platform and Google Maps is the application that runs on the platform, using its compute and storage. Manufacturers need the IIoT platform, and they must either (1) build the IIoT application themselves (which is difficult since manufacturers often don’t have the in-house talent), or (2) buy a prepackaged application from a vendor.
The sub-scale platform problem occurs when, in a market, there are too many platform vendors who can’t make enough money to scale their business due to intense competition and operational execution issues (identified above) and when there aren’t enough applications to actually create value for the customer, the manufacturer. The IIoT market in the United States has faced precisely this problem, especially because digital-platform markets tend toward winner-take-all or oligopoly competition dynamics (eg. iPhone vs. Android; the four major cloud-computing platforms: Amazon, Google, Microsoft, and now Oracle), and platforms make money only if application vendors build on the platform.
BCG, in a 2017 report titled “Who Will Win the IoT Platform Wars,” identified over 400 IoT platforms in the market due to the excitement of the industry at that time. But few of these platforms really grew to any significant scale, with some notable failures (see GE’s attempt above) because of the technical and operational issues. As a result, there were few IIoT application vendors building prepackaged software. There too many platforms they could choose to build on, and the lack of platforms at scale meant there were too many technical challenges that were unresolved. The value of the platform is to solve the underlying technical issues so an application developer doesn’t have to. In the IT world, a software developer doesn’t have to worry about which type of server or networking equipment is in the data center to build a cloud application. The same is true for a software developer on mobile: they don’t have to worry about the specific type of camera lens on the phone when building their app.
As a result, there are few if any IIoT applications at scale (Samsara being a notable exception). For example, there is no packaged software application that a factory own can buy to predict when any robot it chooses to deploy will breakdown today, or for any other type of equipment (of which there are literally thousands) in a factory.
Meanwhile, China’s State Council, in the same 2017 policy document, designed policies to solve the sub-scale platform problem in IIoT:
By 2020, form the industrial internet platform system, supporting the construction of approximately 10 cross-industry, cross-domain platforms, and establishing a number of enterprise-grade platforms that support companies’ digital, internet-enabled, and AI-enabled transformations. Incubate 300,000 industry-specific, scenario-specific industrial applications, and encourage 300,000 enterprises to use industrial internet platforms for research and development design, production manufacturing, operations management, and other business activities. The foundational and supportive role of industrial internet platforms in industrial transformation and upgrading will begin to emerge.
Like most industrial policies in China, the State Council’s high-level policy guidance becomes operationalized in provincial- and city-level policies via funding and other incentives. For example, Jiangsu 江苏 province set a goal of establishing 1,000 “smart” (aka enabled by cloud, AI, advanced connectivity, etc.) factory workshops in 50 provincial-level factories by 2020.
What can the United States learn?
If we’re serious about revitalizing US manufacturing or maintaining leadership in emerging technologies such as AI and quantum computing, here are some things US policymakers should consider:
The free market, while efficient for specific markets, may not optimize for transforming entire sets of industries. The technologies for the industrial internet of things were available in the United States — but due to technical and market-coordination challenges, adoption has lagged behind that of China. AI and quantum are foundational technologies that may require an even greater level of market coordination to overcome operational and technical obstacles compared to that of the industrial internet of things.
Industrial policy needs to move beyond tax incentives, tariffs, and subsidies to make calculated bets on specific technologies, with deep technical expertise incorporated early on in the policy process. For example, in AI, the policy debate has focused exclusively on semiconductor subsidies and export controls — but there is limited if any discussion on how to make the AI data center itself easier to build and operate. High energy costs and energy availability due to the limits of the utility grid are known technical and business challenges to data center capacity today. Ultimately, the total cost of using AI to make predictions, optimize processes, and create value (eg. cost of inference) is not just the cost and efficiency of the chips, but the entire data center stack, including energy costs.
Successful commercialization of a set of technologies creates its own positive feedback loop, which reinforces first-mover advantages. Since China has a significant head start in digitizing its manufacturing base via IIoT technologies, Chinese vendors likely have more real-world data (by deploying more sensors), which enables firms to perfect their machine-learning algorithms, which will further improve manufacturing productivity in China relative to the United States. Robot adoption is a key example: when adjusted for labor costs, China uses 12 times more robots than the United States. This deployment of industrial robots at scale further advantages Chinese manufacturers and the entire technology stack associated with robotics (eg. operating systems for robots, robot supply chain, AI software to control the robots, software integrating robots into production processes, etc.). Recent reports of the Chinese government and enterprises mass-adopting DeepSeek only add urgency for more innovative industrial policies in the United States. Therefore, to achieve policy goals such as restoring US manufacturing or maintaining US leadership in quantum or AI, the United States must support companies to actually buy and use these technologies themselves.
While China may have “won” the initial round of the IoT platform wars, it isn’t too late for the United States, with smart policies and leadership, to win the broader industrial-technical leadership competition with China. While some may object to “picking winners and losers,” without urgent policy action, there may only be losers left to pick from.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
Why is Trump appeasing Russia? What lessons can we learn from the battlefield in Ukraine? How will AI change warfare, and what does America need to do to adapt?
To discuss, we interviewed Shashank Joshi, defense editor at the Economist on a generational run with his Ukraine coverage, and Mike Horowitz, professor at Penn who served as Biden’s US Deputy Assistant Secretary of Defense for force development and emerging capabilities in the Pentagon.
We discuss….
Trump’s pivot to Putin and Ukraine’s chances on the battlefield,
The drone revolution, including how Ukraine has achieved an 80%+ hit rate with low-cost precision systems,
How AI could transform warfare, and whether adversaries would preemptively strike if the US was on the verge of unlocking AGI,
Why Western military bureaucracies are struggling to adapt to innovations in warfare, and what can be done to make the Pentagon dynamic again.
This episode was recorded on Feb. 26, two days before the White House press conference with Zelenskyy, Trump, and JD Vance. Listen now on iTunes, Spotify, YouTube, or your favorite podcast app.
Jordan Schneider: Shashank, it seems you had a lot of fun on Twitter this week?
Shashank Joshi: I was in a swimming pool with my children on holiday in the middle of England and didn’t notice until 18 hours after the fact that the Vice President of the United States had been rage-tweeting at me over my intemperate tweets on the subject of Ukraine. I provoked him into this in much the same way that he believes Ukraine provoked the invasion by Russia.
Jordan Schneider: What does it mean?
Shashank Joshi: It means the Vice President has far too much time on his hands, Jordan.
This is a pretty significant debate. Fundamentally, this was about whether Ukraine is fated to lose. His contention is that Russian advantages in men and weapons or firepower meant that Ukraine’s going to lose no matter what assistance the United States provides.
My argument was that while Ukraine is not doing well — I’m not going to sugarcoat that, I’ve written about this and it’s made me pretty unpopular among many Ukrainians — it’s not true that advantages in manpower and firepower always and everywhere result in decisive wins. Indeed, Russia’s advantage in firepower is much narrower than it was. The artillery advantage has closed. Ukraine’s use of strike drones — which we’ll talk about later — has done fantastic things for their position at the tactical level.
On the manpower side, Russia is still losing somewhere in the region of 1,200-1,300 men killed and wounded every single day. While it can replenish those losses, it can’t do that indefinitely. I’m not saying Vance is completely wrong — I’m just saying he is exaggerating the case that Ukraine has already lost and that nothing can change this.
My great worry is this is driving the Trump administration into a dangerous, lopsided, inadequate deal that is going to be disastrous for Ukraine and disastrous for Europe. I’m worried profoundly about that at this stage.
Michael Horowitz: Quantity generally sets the odds when we think about what the winners and losers are likely to be in a war. Russia has more and will probably always have more. But there are lots of examples in history of smaller armies, especially smaller armies that are better trained or have different concepts of operation or different planning, emerging victorious. Most famously in the 20th century, perhaps Israel’s victory in 1967.
Jordan Schneider: We have three years of data. It’s not like you’re playing this exercise in 2021. You’re doing this exercise in February of 2025. By the way, Mr. Vice President, your government actually has a ton of the cards here to change those odds and change the correlation of forces on the ground, which just makes the argument that this is a tautology so absurd coming from one of the people who is in a position to influence and who has already voted for bills that did influence this conflict.
Shashank Joshi: Wars are also non-linear. You can imagine a war of attrition in which pressures are building up on both sides, but it isn’t simply some mathematical calculation that the side with the greatest attrition fails. It depends on their political cohesion, their underlying economic strength, their defense industrial base, and their social compact.
The argument has been that although Russia feels it has the upper hand — it has been advancing in late 2024 at a pace that is higher than at almost any time since 2022 — there’s no denying that to keep that up, it would have to continue mobilizing men by paying them ever higher salaries and eventually moving to general mobilization in ways that would be politically extremely unpalatable for Vladimir Putin. War is not just a linear process. It’s a really complicated thing that waxes and wanes, and you have to think about it in terms of net assessment.
Michael Horowitz: That’s especially true in protracted wars. I’m teaching about World War I right now to undergraduates at Penn. One of the really striking things about World War I is if you look at the French experience, the German experience, and the Russian experience in particular, given the way that World War I is one of the triggers for the Russian Revolution, how their experience plays out in World War I is in some ways a function of political economy — not just what’s going on on the battlefield, but their economies and the relationship to domestic politics and how it then impacts their ability to stay in and fight.
Jordan Schneider: America has levers on both sides of the political economy of this war. There was a point a few weeks ago when Trump said he was going to tighten the screws on Putin and his economy. The fact that we are throwing up our hands and voting with Putin in the United Nations, saying that they were the aggressor, just retconning this entire past few years is really mind boggling. There was a line in a recent Russia Contingency podcast with Michael Kofman, where he says “The morale in Munich was actually lower than the morale I saw on the front in Ukraine,” which is a sort of absurd concept to grapple with.
Michael Horowitz: If you were to mount a defense here, what I suspect some Trump folks might say is that they believe this strategy will give them more leverage over Russia to cut a better deal. That involves saying things that are very distasteful to the Ukrainians, but they think as a negotiating strategy, that’s more likely to get to a better outcome.
Shashank Joshi: That’s right, Mike. Although they’ve amply shown they are willing to tighten the screws on Zelenskyy. If you were looking at this from the perspective of the Kremlin, would you believe General Keith Kellogg when he says, “If you don’t do a deal, we’re going to ram you with sanctions, batter you with economic weapons"? Or do you listen to Trump’s rhetoric on how we’re going to have a big, beautiful economic relationship with Russia and we’re going to rebuild economic ties, lift sanctions?
You’re going to be led into the belief that the Americans are really unwilling to walk away from the table because the Vice President and others are publicly saying we don’t have any cards, that the Ukrainians are losing, and if we don’t cut a deal now, then Russia has the upper hand. It puts them in a position of desperation.
My big concern is not just that we get a bad deal for Ukraine, it’s that the idea of spheres of influence appeals to Trump, dealing with great men one-on-one, people like Kim Jong Un, Vladimir Putin, Xi Jinping — and that what will be on the table is not just Ukraine, but Europe. Putin will say, “Look, Mr. President, you get your Nobel Peace Prize, we get a ceasefire, we do business together and lift sanctions. And you can make money in Moscow, by the way. Just one tiny little thing, that NATO thing. You don’t like it, I don’t like it. Just roll it back to where it was in 1997, west of Poland. That would be great. You’ll save a ton of money here. I’ve prepared a spreadsheet for you.”
That is the scenario that worries us — a Yalta as much as a Munich.
Jordan Schneider: We have a show coming out with Sergey Radchenko where we dove pretty deep into Churchill’s back-of-a-cocktail-napkin split. At least Churchill was ashamed.
It’s so wild thinking about the historical echoes here. I was trying to come up with comparisons, but the only ones I could do were hypotheticals. Like McClellan winning in 1864, or — I mean, Wendell Willkie was actually an interventionist. There was some Labor candidate that the Nazis were trying to support in the Democratic Party in 1940, but he never made it past first base. Has there ever been a leadership change that shifted a great power conflict this dramatically?
Shashank Joshi: From the Russian perspective, that’s Gorbachev. Putin would look back at glasnost, perestroika, and Gorbachev at the Reykjavik summit as moments where a reformist Soviet leader sold the house to the Americans and threw in the towel.
Michael Horowitz: You also see lots of wars end with leader change, with leadership transitions, when wars are going poorly for countries and you have leaders that are all in and have gambled for resurrection. If you think about the research of someone like Hein Goemans back in the day, then you have to have a leadership transition in some ways to end wars in some cases if leaders are sort of all in on fighting.
Jordan Schneider: The Gorbachev-Trump comparison is a really apt one because it really is like a true conceptual shift in the understanding of your country’s domestic organization as well as role in the world. Gorbachev, for all his faults, at least had this universalist vision of peace, trying to integrate in Europe — he wanted to join NATO at one point. But going from that to whatever this 19th century mercantilism vision is, is really wild to contemplate.
Shashank Joshi: The other thing to remember is Gorbachev’s reforms eventually undid the Soviet empire. They undid its alliances and shattered them. In the American case, the American alliance system is not like the Soviet empire. France and the UK are not the Warsaw Pact. We bring something considerably more to the table. It’s a voluntary alliance. It’s a technological, cultural alliance. These are different things.
I worry sometimes that this administration or some people within it — certainly not everybody — views allies just as blood-sucking burdens. What they don’t fully grasp is how much America has to lose here. I want to say a word on this because Munich — and I heard this again — the FT reported recently that some Trump administration official is pushing to kick Canada out of the Five Eyes signals intelligence-sharing pact.
Now okay, the Americans provide the bulk of signals intelligence to allies. There’s no surprise about that. But if you lost the 25% provided by non-US allies, it will cost the US a hell of a lot more to get a lot less. It will lose coverage in places like Cyprus, in the South Pacific, all kinds of things in the high north, in the Arctic in the Canadian case. This administration just doesn’t understand that in the slightest.
Michael Horowitz: Traditionally what we’ve seen is regardless of what political hostility looks like, things like intelligence sharing in something like the Five Eyes context continues — in some ways the professionals continue doing their jobs. If you see a disruption in that context, that would obviously be a big deal.
Jordan Schneider: Just staying on the Warsaw Pact versus NATO in 2025 today, America plus its allies accounted for nearly 70% of global GDP during the Cold War. The economic outflows that were needed to sustain Soviet satellites eventually bankrupted the USSR. America isn’t facing anything resembling that situation by stationing 10,000 people in Poland and South Korea.
Michael Horowitz: We are in a competition of coalitions with China, and it is through the coalition that we believe we can sustain technological superiority, economic superiority, military power, et cetera. Look at something like semiconductors and the role that the Netherlands plays in those supply chains, that Japan plays in those supply chains. There are interconnections here. We have thought that we will win because we have the better coalition.
Shashank Joshi: That’s an interesting question to ask more conceptually — does this administration want a rebalancing of its alliances or does it want a decoupling? You could put it in terms of de-risking and decoupling if you want to echo the China debate here. Does it simply want more European burden-sharing? But fundamentally the US will still maintain a presence in Europe, underwrite European security, and provide strategic nuclear weapons as a backstop. That is what many governments are trying to tell themselves.
The more radical prospect is that whilst there are some people who envision that outcome — Marco Rubio, Mike Waltz (the National Security Advisor), and John Ratcliffe (the head of the CIA) — the President and many of the people around him view things in considerably more radical terms. It’s more of a Maoist cultural revolution than a kind of “I’m Eisenhower telling the Europeans to spend more.”
Jordan Schneider: There’s this quote from Marco Rubio that’s really stuck with me from a 2015 Evan Osnos profile where he talks about how he has not only read but is currently rereading The Last Lion, which is this truly epic three-part series. The middle book alone is most famous, which is what Rubio was referring to, where Churchill saw the Nazis coming when no one else did and did everything he could in the 30s to wake the world up and prepare the UK to fight.
Rubio is referring to this moment by comparing it to how he stood up to the Obama administration when they were trying to do the JCPOA nuclear deal with Iran. To go from that to having to sit on TV and blame Ukraine for starting the war, I think is just the level of cravenness. There are different orders and degrees of magnitude.
Secretary of State Marco Rubio looking very uncomfortable, February 28th, 2025. Source.
Shashank Joshi: You have to think about this not in terms of a normal administration in which people do the jobs assigned to them by their bureaucratic standing. You have to think about it like the Kremlin, where you have power verticals, or an Arab dictatorship where you have different people reporting up to the president. Think of this like in Russia, where you have Sergey Naryshkin, the head of the Foreign Intelligence Service, who may say one crazy batshit thing, but actually has no authority to say it. In which Nikolai Patrushev may say another thing, in which Sergey Lavrov may lay down red lines, but they have no real meaning because there’s a sense of detachment from the brain, the power center itself. Ultimately, it’ll still be Putin who makes the call. I think it’s a category error if we try to think about this administration as a normal system of American federal government.
Michael Horowitz: I will say, I can’t believe I’m now going to say this, but let me push back and say that there’s a lot of uncertainty about what the Trump administration wants to accomplish here, given the way they have embraced the notion that Trump is a master negotiator. To be professorial about it, in a Thomas Schelling “threat that leaves something to chance” way, or like madman theory kind of way, they think that there’s a lot of upside here from a bargaining perspective.
Most of Trump’s national security team is not yet in place. We just had a hearing for the Deputy Secretary of Defense yesterday. Elbridge Colby, who’s the nominee for undersecretary, has a hearing coming up, I think either next week or the following week. So a lot of the team is still getting in place.
Jordan Schneider: The thing about Trump 1.0 is there weren’t wars like this. You had two years of sort of normal people who were basically able to stop Trump from doing the craziest stuff. Then the COVID year was kind of a wash. But Trump 2.0 matters a lot more, it’s fair to say, over the coming four years than it did 2016-2020.
Shashank Joshi: It’s much more radical. In the first term, John Ratcliffe had his nomination pulled as DNI because he was viewed as inexperienced and not up to the job. Today, John Ratcliffe looks like Dean Acheson compared to the people being put into place. We have to pause and make sure that we recognize the radicalism of what is being put into place around us.
When you look at the sober-minded people who thought about foreign policy — and I include amongst this people I may disagree with, like Elbridge Colby, who will be probably the Pentagon’s next policy chief — what is the likely bureaucratic institutional political strength they will bring to bear when up against those with a far thinner history of thinking about foreign policy questions?
Jordan Schneider: I haven’t done a Trump-China policy show because I don’t think we have enough data points yet. But what, if anything, from the past few weeks of how he’s thinking and talking about Russia and Ukraine, is it reasonable to extrapolate when thinking about Asia?
Shashank Joshi: Two quick things. One is I see significant levels of concern among Asian allies. The dominant mood is not, “Oh, it’s fine, they’re going to just pull a bunch of stuff from Europe, stick it into Asia and it’ll be a great rebalancing.”
Number two, I think this is important: there is a strong current of opinion that views a potential rapprochement with Russia as being a wedge issue to drive between Russia and China, the so-called reverse Kissinger. Jordan, you know much more about China than I do. I’m not going to comment further on that, but I will say I believe it is an idea that is guiding and shaping and influencing current thinking on the scope of a US-Russia deal.
Michael Horowitz: You certainly have a cast of officials who are pretty hawkish on China, which will be a continuation in some ways of the last administration and the first Trump administration. I think the wild card will be the preferences of the president. There was a New York Times article a few days ago that talked about Trump’s desire for a grand bargain with China — his desire to do personal face-to-face diplomacy with Xi as a potential way to obtain a deal.
Trump hosts Xi Jinping at Mar-a-Lago in 2017. Source.
Now I think the reality is that every American president that has tried to do that kind of deal, whether in person or not over the last decade, has found that there are essentially irreconcilable differences. There’s a reason why there is US-China strategic competition and why that has been the dominant issue in some ways of the last several years and probably will be over the next generation. But Trump may wish to give it a shot — and it sounds like, at least from that article, that he might.
Jordan Schneider: We’ve also had every administration in the 21st century try to start their term by trying to reset relations with Russia. “Stable and predictable relationship” was Biden’s line. Maybe this stuff is just a blip, but I think Shashank’s right. We’re in really uncharted territory.
Paid subscribers get access to the rest of the conversation, where we discuss…
AI as a general-purpose technology with both direct and indirect impacts on national power,
Whether AGI will cause instant or continuous breakthroughs in military innovation,
The military applications of AI already unfolding in Ukraine, including intelligence, object recognition, and decision support,
AI’s potential to enable material science breakthroughs for new weapons systems,
Evolution of drone capabilities in Ukraine and “precise mass” as a new era of warfare,
How China’s dependence on TSMC impacts the likelihood of a Taiwan invasion,
Whether AGI development increases the probability of a preemptive strike on the US,
How defense writers and analysts help shape policy and build bureaucratic coalitions,
Ukraine as a real-world laboratory for testing theories about warfare, and what that means for Taiwan’s defense.
Jordan Schneider: Let’s talk about the future of war. There is this fascinating tension that is playing out in the newly national security-curious community in Silicon Valley where corporate leaders like Dario Amodei and Alex Wang, both esteemed former ChinaTalk guests, talk about AGI as this Manhattan Project-type moment where war will never be the same after one nation achieves it. What’s your take on that, Mike?
I’d like to spotlight the newest NSF directorate, Technology, Innovation and Partnerships (TIP) created by the CHIPS & Science Act, that has been particularly hard-hit by DOGE. The idea was to supplement the world-class basic research that NSF does with more use-inspired and translational research with higher technology readiness levels. I’ve been following this directorate since its creation, recorded a panicked emergency pod when for a hot minute Senate Commerce almost killed it, and have been really impressed with its work so far.
TIP helped stand up NAIRR, has done a fanstastic job helping catalyze regional innovative hubs, and is the only org I’ve seen in government actually be strategic about workforce development. My personal favorite its new APTO program, which is creating the data and intellectual substrate necessary to really do smart S&T and industrial policy. For more of what TIP has been up to, check out their Director’s annual letter here. I’d also encourage DOGE to have a read of the TIP’s roadmap for the next few years and try to spot stuff that America doesn’t need.
The NSF is not perfect. IFP has some excellent proposals on how to incorporate novel funding strategies like lotteries that need faster adoption. But IFP also recently wrote up how the NSF showed its mettle, and was able to move faster than the NIH for COVID-related grants. TIP in particular has collected some of NSF’s most forward-thinking talent and is experimenting with novel programs and funding strategies faster than anyone else in the NSF mothership.
American basic research is our golden goose and the envy of the world, building the basis for scientific innovations that make us richer, live longer, and make us more powerful. Our universities attract the best minds in the world which is an enormous boon to the country, and absent radical intervention will continue to do so. While the NSF could use reform, we are criminally underfunding R&D already, and firing the most forward-thinking junior staff in the directorate singled out by national security heavyweights as critical to competing with China is an error this administration should correct.
Try Picking on Someone Your Own Size
DOGE should really try taking on some government programs that aren’t already running lean, creating the future, preventing pandemics and saving lives. The real discretionary bloat isn’t malaria bednets and fundamental physics research but F-35s and carriers. A real push at a few deadweight DoD programs could deliver way more savings than whatever you can squeeze from NSF and USAID and likely make for a more effective force.
The only way the DoD was really going to change was through major budget cuts — something that forced people’s hands into new ways of working, into true prioritization, into processes that took less time because they were less burdened by the trappings that come with enormous budgets. I began my comment with an apology to the senior Air Force official sitting next to me, a caveat that I meant no disrespect, and wasn’t arguing for less military might — in fact, what I wanted was a more capable military. To my surprise, he piled on. “She’s right,” he said. “But it has to be much deeper than anything we’ve seen before. We had to cut during the last sequestration, and it was around 15% off the top of everything, which doesn’t force meaningful choices. It needs to be like half.”
To get at wasteful DoD programs and acquisitions regulations this administration would have to do the hard work of wooing Congresspeople into taking votes that would more substantially impact their districts. I hope that Trump 2.0’s staff has the stomach and topcover for this sort of work that could yield real long-term dividends for the country, not just grabbing the lowest hanging political fruit which really even have long term fiscal relevance like cutting probationary employees, foreign aid, and basic R&D.
From a ChinaTalk episode coming out on Monday with Mike Horowitz, former Biden DoD official, and The Economist’s Shashank Joshi:
Jordan Schneider: And I think this is like one of the many shames of the Trump imperial presidency. He has enough control of Congress to do this well and could even get some Dem votes for real defense reform!
Mike Horowitz: Let me muster a point of optimism here. If you look at Hegseth's testimony, his discussion of defense innovation is very coherent. He has takes that are not structurally dissimilar to the ones that we have been making.
There is a potential opportunity here for the Trump administration to push harder and faster on precise mass capabilities, on AI integration, and on acquisition reform in the defense sector. Because the president right now seems to have a strong hand with regard to Congress. Whether the president's willing to use political capital for those purposes is not clear. How the politics of that will play out is unclear. But if the Trump administration does all the things that it says it wants to do from a defense innovation perspective, that may not be a bad thing!
Shashank Joshi: My concern is also that you have people who are good at radicalizing and disrupting many businesses and sectors and fields of life. But the skills that are required to do that are different to the skills in a bureaucracy like this. Because, just because you were able to navigate the car sector and the rocket sector, doesn't mean you know how to cajole, persuade, and massage the ego of a know-nothing congressman who knows nothing about this and who simply cares that you build the attributable mass in his state, however stupid an idea that is, and who wants you to sign off on the 20 million dollars.
I worry that they will either break everything, and I'm afraid what I'm seeing DOGE do right now with a level of recklessness and abandon is worrying to me as an ally of the United States from a country that is an ally, but also that they will just not have the political nous [British for common sense] to navigate these things to make it happen. Just because Trump controls Congress and has sway over Congress doesn't mean that the pork barrel politics of this at the granular level fundamentally change. You need operatives, congressional political operatives. A tech bro may have many virtues and skills, but that isn't necessarily one of them.
Here’s to hoping! Howabout a Washington quote to send us off, from a 1775 letter sent to General Schuyler: “Animated with the Goodness of our Cause, and the best Wishes of your Countrymen, I am sure you will not let Difficulties not insuperable damp your ardour. Perseverance and Spirit have done Wonders in all ages.”
The AI Action Summit, which closed just over two weeks ago in Paris, will be remembered as a historically important gathering — though not how many of its organizers, attendees, and contributors anticipated. Rather than cementing AI safety as a priority for transnational collaboration, it turned into a memorial service for the safety era.
This Summit’s lasting moments, however, came not from the success of “open, multi-stakeholder and inclusive approach[es]” on safety championed by the official declaration from the event, but instead dramatic declarations of national primacy unshackled by safety concerns. Vice President JD Vance’s speech made little accommodation for either safety or internationalism, declaring that the United States was “the leader in AI and our administration plans to keep it that way,” and that he was not here to talk about AI safety but instead “AI opportunity.” Macron touted a massive €110 billion fund to back AI projects in France, and the United States and United Kingdom declined to sign the Summit’s declaration language. A wildcat “Paris Declaration on Artificial Intelligence” backed by private industry hit the Summit for failing to back a “strong, clear-eyed, and Western-led international order for AI.”
A sense of stuckness prevailed in the side conversations and events taking place throughout Paris. At the AI Security Forum, a slow carousel of speakers ran through very much the same tropes and ideas that had dominated the discourse for years. Shakeel Hashim captured a feeling widely held — that the Summit was a “pantomime of progress” rather than the genuine article.
This isn’t just a vibe. The “AI safety community” has always nurtured a shared, but often unspoken, agreement that public-minded technical expertise and international cooperation were the most promising pathways to promote good global governance of the technology.
But the safety community made a historically bad bet. The wheels were already coming off multistakeholder, international governance in the world at large even as the safety community began to invest in it seriously in the mid-2010s. Resurgent nationalism, great-power competition, and the fecklessness of international institutions have limited options for global governance across many domains, and AI has been just another one of the casualties. This isn’t just about Trump winning: these changes in the international system are structural, and the domestic shifts in places like France and the UK would have led to a very similar result even if Harris had pulled it out last year.
The safety community was also profligate in the use of its attention and social capital. The political influence of fair-minded technical experts turned out to be a rapidly depleting resource, wasted away as one “high-profile letter from very concerned scientists” and “dramatic demo of hypothetical model threat” followed another to little effect.
Against such a backdrop, it’s no wonder that AI safety in 2025 feels ever more like pantomime. We’re still frantically pulling the same levers, even as the whole constellation of forces that move nations in general and technology policy in particular have rearranged.
We need to be asking hard questions. What are historical models for technological safety and stability in a world of fierce, unrestricted nationalism? What happens when scientific evaluation has lost its ability to persuade the policymaker? How do you slow down or stop a technological race-in-progress?
The real intellectual work is now rebuilding a theory for safety that takes these uncomfortable realities into account and builds as best it can around them.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
Can economic warfare really work? What can we learn from the 21st century historical record of American sanctions policy?
To find out, we interviewed Eddie Fishman, a former civil servant at the Department of State and an Adjunct Professor at Columbia. His new book, Chokepoints: American Power in the Age of Economic Warfare, is a gripping history of the past 20 years of American sanctions policy.
In this show, we’ll talk about…
The evolution of U.S. sanctions policy, from Iraq and Cuba to Iran and Russia,
How Reagan’s deal with the Saudis turned the dollar into an economic chokepoint,
The incredible success of sanctions against Iran, and how that playbook could have been used to punish Russia,
Historical lessons in enforcement that are relevant for export controls on China today,
The role of great civil servants like Stuart Levey, Daleep Singh, Victoria Nuland, and Matt Pottinger in building state power,
Institutional challenges for economic warfare and the consequences of failure to reform,
Strategies for writing groundbreaking books about modern history.
Jordan Schneider: Let’s start with the Bosphorus. How does this little corner of our beautiful planet explain the evolution of sanctions?
Edward Fishman: The Bosphorus is the epitome of a maritime chokepoint. It is a narrow strait between the Black Sea and the Mediterranean Sea. Throughout history, maritime chokepoints like the Bosphorus have been critical for strategic power. Sparta was able to win the Peloponnesian War because they won a battle around the Bosphorus and blockaded it, ultimately starving the Athenians into submission. Athens had relied on the flow of grain through the Bosphorus to feed its population — that was really the whole purpose of ancient Athens’ maritime empire.
Historically, these chokepoints have been geographic features. But now, as a result of globalization, there are chokepoints in the global economy that are not geographic — the most critical of which is the U.S. dollar. This is why the book is called Chokepoints.
For thousands of years throughout history, the only way to block a maritime chokepoint like the Bosphorus was a physical naval blockade. What’s changed is that in the wake of hyperglobalization in the 1990s, the U.S. acquired the ability to block chokepoints like the Bosphorus just by weaponizing its control of the U.S. dollar.
Today, the director of OFAC, the unit at the Treasury Department that oversees sanctions policy, can sign a few documents in her office and blockade a chokepoint like the Bosphorus. This actually happened on December 5, 2022, when the G7 oil price cap went into effect. The Bosphorus was backed up with dozens of oil tankers, because Turkish maritime officials were so nervous about violating the terms of the price cap that they didn’t want the ships to cross. It took OFAC days of very intensive diplomacy with Turkish authorities to persuade them to allow the ships to cross.
Source: Chokepoints, pg 2
Jordan Schneider: You open this book with some wild contrast. Historically, you needed triremes. Now, all you need is a piece of paper from the Treasury Department to clog up the strait in Turkey halfway around the world.
Like you, Eddie, I was a sanctions nerd in college. I wrote my thesis about the origins of the UN and did papers on sanctions policy. I remember very vividly reading this literature arguing that sanctions are useless and don’t have any big impact. There was this great quote from George W. Bush in your book where at some point in the 2000s, he said, “We’ve sanctioned ourselves out of any influence” when it came to Iran’s nuclear program. You put the spotlight on one civil servant who takes that as a challenge and through ingenuity, creativity, and a whole lot of elbow grease, is able to discover and leverage a whole new lens of American power. Let’s briefly tell the story of American sanctions pre-Stuart Levey before we discuss Iran’s nuclear program.
Edward Fishman: When Stuart Levey came in as the Treasury Department’s first undersecretary of terrorism and financial intelligence in 2004, the most recent big case of sanctions that the U.S. had was a 13-year sanctions campaign against Iraq from 1990, when Saddam originally invaded Kuwait, until 2003, when George W. Bush launches the invasion of Iraq. That embargo required full UN backing and was implemented by a 13-year naval blockade. You had literally a multinational naval force parked outside of Iraqi ports inspecting every single oil shipment going in and out of Iraq.
The lesson from this situation was that sanctions didn’t work — Saddam didn’t come to heel. He seemed to be just as aggressive, if not more so. Over time, this embargo wound up leading not only to humanitarian problems in Iraq, which are very well documented, but also significant corruption. Saddam was siphoning away oil money under the nose of the UN.
By the time Levey comes in, sanctions had been seen as something that had been tried and failed against Iraq, and in fact had paved the way for the U.S. invasion of Iraq. In many ways, the 2003 invasion of Iraq was a direct result of the perception that sanctions had failed.
When Levey started working on the Iran problem around 2004, the prospect of even doing an Iraq-style sanctions campaign against Iran was off the table because there was no way to get the UN Security Council to agree to that at the time. Bush’s comment about having sanctioned ourselves out of influence with Iran was a result of the fact that without the UN, the U.S. thought that the only type of sanctions we could impose were primary sanctions, like an embargo where U.S. companies can’t buy things from Iran or trade with Iran. The only issue is we had had an embargo in place since the mid-90s, so there wasn’t any trade to speak of between the U.S. and Iran. The two avenues of sanctions were closed off — sanctions through the UN had been discredited by the 90s, and the other, primary sanctions on Iran, had already been maxed out and had been for a decade by then.
Jordan Schneider: The other seminal piece of sanctions in American 20th-century history is the embargo on Cuba. That is the same story — we cut off trade with this country, yet Castro’s still there in 2004, some 50-odd years later. It’s interesting — if you go back even further, there was this real hope after World War II where the UN at one point was even going to have its own air force. The idea was that sanctions were going to be this incredible tool to deter bad actions by different actors around the world because the U.S. and the Soviet Union were friends and we would all police the planet in a happy-go-lucky way. That was not how the Cold War ended up working out.
In 2004, Stuart Levey started to understand that he can leverage the dollar’s role in global financial flows. Eddie, can you tell the story of how the U.S. dollar became globalized in this way?
Edward Fishman: Bretton Woods, the conference that set the rules of the road for the post-World War II economy, happened in 1944. It put the U.S. dollar at the center of the global economy and established the dollar as the global reserve currency. It made the dollar as good as gold — the dollar is convertible for a fixed rate of $35 per ounce of gold.
At the same time, it explicitly prioritized the real economy and trade over finance. John Maynard Keynes, who was one of the architects of the Bretton Woods system, said that capital controls were a very important part of the system. For the first 30 years of this new global economy that emerged after World War II, you had the dollar at the center of the world economy, but it wasn’t a particularly financialized world economy. Most states had pretty significant capital controls, and banking was a very nationalized and, in some ways, even just a regionalized type of business.
By 1971, the U.S. dollar had been losing its value for quite some time and we were running significant deficits because of the war in Vietnam. Ironically, this is when Richard Nixon unilaterally took the dollar off of the gold peg. The dollar was still at the center of the world economy, but it was no longer tethered to gold. Exchange rates were now set by the market instead of by government fiat.
In the years after that, the capital controls of the Bretton Woods system fully erode and the dollar winds up becoming even more integral to the world economy as we see financialization take off from the ’70s through the Clinton era. You get to the point where we have a foreign exchange market that is turning over seven or eight trillion dollars every single day, which is by far the largest of all financial markets.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
Jordan Schneider: How did oil come to be traded in U.S. dollars?
Edward Fishman: The dollar’s role in trading oil is arguably the most important chokepoint for a number of the key sanctions campaigns of the 21st century.
After World War II, the U.S. was a large oil producer and a big exporter. The 1973 Arab oil embargo shifted our perspective, and the U.S. realized just how vulnerable it was to being cut off from Middle Eastern oil.
In 1974, Richard Nixon — who was wallowing under the political pressure of the Watergate scandal and massive deficits that we had no reasonable way of plugging — sent his treasury secretary, Bill Simon, to make a deal. Simon was a former bond trader, a New Jerseyite, a chain smoker...
Jordan Schneider: A chain-smoking New Jersey native, described by a peer as, “far to the right of Genghis Khan.”
Edward Fishman: He’s a really colorful figure. The book includes a photo of him testifying before Congress with a giant plume of smoke around him.
Bill Simon tried to think about how to plug these deficits using his financial background as a bond trader. He proposed cutting a deal with the Saudis such that, not only do they agree to keep pricing oil in dollars into perpetuity, but they actually take the dollars they earn from selling oil and reinvest them in U.S. government debt — they basically plug our deficit with the money that the U.S. is paying them for oil. He wound up taking a flight to Jeddah in the summer of 1974 — getting copiously drunk en route.
Source: Chokepoints, pg. 30
The deal worked. He cut a deal with the Saudis in which they agree to recycle their petrodollars into U.S. Treasuries. This agreement largely still exists to this day. Oil, by and large, is priced in dollars no matter who’s buying it or selling it.
Chokepoints in the global economy are typically formed by the private sector. They kind of develop naturally as businesses evolve. However, there are important moments when government intervention becomes critical.
Simon’s original deal in 1974 solidified the petrodollar, but then a few years later, as the dollar continued to slide in value, oil exporters and OPEC started getting upset because the weakening dollar was in turn reducing the real value of their oil earnings. Jimmy Carter’s Treasury Secretary, Michael Blumenthal, actually went back to Saudi Arabia and cut a new deal in which he agreed to give Saudi Arabia more voting shares at the IMF in exchange for Saudi continuing to price oil in dollars.
Jordan Schneider: Why did the Saudis even cut the deal in the first place?
Edward Fishman: The Saudis got two things. First, they got access to US military equipment, which was pretty beneficial to them. Second, which I think is more of a direct part of this deal and one that’s more easily provable through historical documents, the Saudis were able to buy U.S. government debt in secret outside of the normal auctions. Instead of participating in the public auctions for U.S. Treasuries, they had their own side deal where they could buy Treasuries. That was a big benefit to them because they were able to lock in prices and also do so without facing potential political opprobrium.
Jordan Schneider: That’s crazy.
Edward Fishman: It’s a remarkable turning point in the financial and economic history of the 20th century. There was a real shot that oil could have been priced against a basket of currencies, which in some ways makes more sense. For these countries in the Middle East and OPEC members, their entire economy basically depends on generating oil revenue. If you want stability and predictability, you don’t want to take exchange rate risk. But people like Bill Simon and Michael Blumenthal intervened and were able to get the dollar enshrined as the key part of the oil market.
The Iran Sanctions Formula and JCPOA Diplomacy
Jordan Schneider: Let’s talk about 2006, when Stuart Levey was trying to figure out how to make sanctions work against Iran. Can you explain his light bulb moment during the January 2006 trip to Bahrain?
Edward Fishman: Levey realized other countries hadn’t stopped doing business with Iran — only the U.S. had, and that’s why the sanctions weren’t working. But he realized that he could use access to the dollar as a lever to pressure foreign banks.
Typically, when you’re trying to get other countries on board for sanctions, you would go negotiate with their foreign ministry and say, “We think what Iran’s doing is bad. You should impose your own sanctions on Iran.” That was the paradigm before 2006. What Levey realizes is that he can go directly to the CEOs of foreign banks, bringing declassified intelligence demonstrating how Iran uses their banks to finance their nuclear program, and funnel money to terrorist proxies like Hamas and Hezbollah. To start, he could just present the facts and potential reputational concerns would often persuade these banks to exit Iran. In more extreme circumstances, when banks wouldn’t go along with him, he could threaten their access to the dollar to try to get them out of Iran.
What Levey really pioneered was the direct diplomacy between him as a Treasury official and his team at the Treasury Department with bank CEOs. You might ask, how did Stuart Levey get meetings with CEOs of banks all around the world? He was lucky — right when he had this epiphany, Hank Paulson, who had been the CEO of Goldman Sachs, came in as Treasury Secretary. Paulson is arguably the most well-connected banker in the world at the time. Hank winds up opening a lot of doors for Stuart and getting him meetings with ultimately more than 100 of the key banking CEOs around the world.
Jordan Schneider: Interestingly, you have to convince all the banks to get on board, because even the slightest institutional leakage would allow Iran to sell as much oil as they want.
How did Levey and his team go about convincing the Russians, the random Chinese banks, the Azerbaijani banks, and all of these other banks?
Edward Fishman: What Levey succeeds at doing between 2006 and 2010 is getting the big name-brand global banks to exit Iran. By and large, there are a few stragglers like BNP Paribas. Most of the big main global banks are out of Iran by 2010, though there are still some banks in places like the UAE, Turkey, and other countries doing business with Iran.
What winds up happening at that time is Congress, which has very little faith in Barack Obama’s willingness to come down hard on Iran — namely because Obama had very explicitly run for president in 2008 saying he wanted diplomacy. He even exchanged letters with Ayatollah Khamenei.
Even Iran hawks that are on the Democratic side of the aisle, like Bob Menendez, don’t really have much confidence that Obama is going to be tough on Iran. Democrats and Republicans basically form almost a coalition against the Obama administration on Iran sanctions and wind up passing progressively harsher sanctions legislation.
The key part of these sanctions laws, the first one called CISADA (the Comprehensive Iran Sanctions Accountability and Divestment Act of 2010), is that they require the Obama administration to impose what’s called secondary sanctions. That’s not sanctions directly on Iran, but sanctions on Iran’s business partners — for instance, the UAE or Turkish bank that I mentioned before.
Iran's Foreign Minister Javad Zarif meeting with Secretary of State John Kerry in July 2014. Source.
Levey was a Bush appointee retained by the Obama administration (he’s one of only two very senior officials, along with Bob Gates, who’s kept on). He uses this law with the mandatory secondary sanctions as a significant cudgel. He goes to places like Dubai and talks to banks saying, “Look, if you don’t get out of Iran, I will be forced by American law to impose sanctions on you. You will lose access to the dollar and all of your assets will be frozen.” That threat is very significant. When the choice is between Iran and the United States dollar, it’s a pretty easy choice for most banks around the world.
Secondary sanctions had been tried before in the mid-90s, but the U.S. effectively wound up blinking and not imposing secondary sanctions on Total, the French oil company that had been investing in Iran’s oil sector. Even the George W. Bush administration decided not to impose secondary sanctions. This tool was very controversial. You can imagine it didn’t go down well with other countries. If you’re an American diplomat and you go meet with one of your counterparts abroad and say, “Sorry, we have to sanction your biggest bank if they don’t stop doing business with Iran” — that just feels like mafia diplomacy, not something that goes down very easily.
One of the virtues of Obama being so beloved around the world was the success of sanctions on Iran. Obama built international consensus that Iran’s nuclear program was a problem.
Jordan Schneider: We also had multilateral sanctions from the UN alongside U.S. action. What did that end up doing for the Obama psyche and the global push to limit Iran’s oil revenue?
Edward Fishman: Obama successfully got a major UN Security Council resolution done in the summer of 2010, right alongside when CISADA, the secondary sanctions law, passed Congress.
Jordan Schneider: In the Medvedev era, mind you.
Edward Fishman: Yes, exactly. Historical contingency matters — the fact that Medvedev was president of Russia at the time meant that Russia didn’t veto UN Security Council Resolution 1929. In retrospect, the benefit of that resolution wasn’t so much the specific sanctions it imposed on Iran. Rather, it explicitly drew connections between Iran’s banking system and energy sector with its nuclear program. This meant when Obama officials traveled the world to tell foreign banks and their governments that they’d be forced to impose sanctions if they didn’t stop doing business with Iran, they could credibly say they were just complying with UN Security Council Resolution 1929 and that international law was on the side of the United States. The legitimacy that Obama’s sanctions campaign derived from the UN was ultimately very significant.
Jordan Schneider: Iran was completely unprepared for this. They literally took out ads in newspapers in Austria to beg for help financing their nuclear program.
Austria Bank reportedly had no idea that this account was being used to help finance Iranian nuclear reactors — until Stuart Levey presented them with a copy of the advertisement above. Source: Chokepoints
Edward Fishman: Exactly. This speaks to assumptions about how the global economy worked at the time. People just trusted that banking networks wouldn’t be weaponized. Iran really thought that they could publicly advertise these fundraising activities with no issue. Foreign banks weren’t aware of what Iran was doing and weren’t particularly worried about being penalized for it. They probably viewed sanctions as something that were unlikely to happen to them — and if they did happen, they could just be chalked up as a cost of doing business.
Jordan Schneider: Let’s talk about the penalties. One of the remarkable accomplishments of the Treasury Department, which the export controls regime on China over the past few years hasn’t been able to do, was the billion-dollar fines thrown on violators — $2 billion on HSBC, and almost $10 billion on BNP Paribas. How did this work?
Edward Fishman: This is a very important part of the story and one that often goes unnoticed. It’s not that sanctions didn’t exist before this period in the early part of the 21st century — it’s that the cost of violating them wasn’t particularly high.
One of the most important strategic legacies of the campaign against Iran pioneered by Stuart Levey is conscripting banks to be frontline infantry of American economic wars. This wasn’t because banks decided that this was morally righteous, it was because they realized that violating sanctions was existentially dangerous for their businesses.
Between 2010 and 2014, Standard Chartered wound up getting fined about a billion dollars, HSBC was fined $2 billion, and BNP Paribas was fined $9 billion. In each case, the New York Department of Financial Services actually threatened to withdraw banking licenses from each of those banks, which would eliminate their ability to do business in the United States. That was a sword of Damocles hanging over these banks — U.S. law enforcement probably could have extracted even bigger fines.
We’re still living with that legacy today. The reason that financial sanctions in particular are so powerful is a confluence of two factors.
The dollar is essential to international commerce. Trying to do business across borders without access to the dollar is like trying to travel without a passport.
The U.S. actually can weaponize the systemic significance of the dollar because banks are afraid of going against American government dictates.
Jordan Schneider: The political economy of it is also different than whacking Nvidia or Synopsys, becauce those three banks are foreign. It is one thing to threaten with extinction some hoity-toity French bank that sponsors the French Open and has been doing business with Iran forever. It’s another to threaten a major contributor to America’s national competitiveness, employment, and growth.
Compare the death sentence of being cut off from the New York Federal Reserve versus mere fines in the case of export controls. With Huawei, there were some cases where they threatened to put executives in jail. Over the past few years, the types of companies that the Biden administration has gone after have often been random Russians in Brooklyn smuggling chips into Russia and China. Whereas the Obama administration was trying to put teeth behind big economic warfare efforts by throwing down billion-dollar fines.
Edward Fishman: Is it possible to conscript tech companies in the same way that banks are conscripted? My own view is yes. If the fines were harsh enough and if the enforcement were strong enough — because the other fact we haven’t talked about is it wasn’t just fines for these banks, it was also independent monitors. The Justice Department sent in people to oversee compliance reforms for several years thereafter.
It is possible, though politically challenging, on one hand to be subsidizing American semiconductor companies to the tune of 50-plus billion dollars, and then on the other to say we’re going to take that money back because you’re violating export controls. It is possible.
One thing I would mention though is that with the BNP fine and the HSBC fine, those took many years to come to fruition. These were years and years of bad behavior that then eventually led to giant fines. It is possible that someone right now at the Justice Department is working away at a major export control violation case that we’ll learn about maybe in a couple of years.
Jordan Schneider: You mentioned “Mafia diplomacy” as a sort of derogatory term for sanctions tactics. There are a lot of moments in this story where gentlemanliness appears to be very important to Obama.
After the invasion of Crimea, around the Maidan revolution, Obama had a call with Putin where he warned that “Moscow’s actions would negatively impact Russia’s standing in the international community.” Putin’s response was basically like, “I don’t know, man, it’s hard to take you seriously.”
Why was Obama’s demeanor so helpful in the case of Iran?
Edward Fishman: Obama was very attuned to international law, or as you put it, gentlemanliness. You could argue he was very lawyerly in his approach. With respect to the Iran sanctions, I think it actually wound up being helpful because the secondary sanctions against Iran were beyond anyone’s imagination.
We haven’t talked yet about the oil sanctions, which were put in place in 2012. The U.S. successfully reduced Iran’s oil exports from 2½ million barrels a day to 1 million barrels a day over about a year. This is explicitly a unilateral U.S. sanction.
Would that have worked as well had Obama not been as attuned to diplomacy and invocations of international law? I’m not so sure. You may have seen more challenges from places like China and India and maybe more obstinance. I do think it was helpful in some regards.
Looking at all the various examples of economic warfare that I talk about in the book, this is in some ways the most remarkable because of how unlikely it is to succeed. But it works.
One big exception from the financial sanctions during the Stuart Levey era is the Central Bank of Iran. The Central Bank of Iran is not under sanctions because it’s the repository for all of Iran’s oil revenues. The Obama administration was really nervous that if they sanction the Central Bank of Iran, other countries won’t be able to pay Iran for its oil. All of a sudden you’ll have all of Iran’s oil go off the market overnight, you’ll have a giant spike in oil prices, and everyone will be in a world of hurt.
Senator Bob Menendez, who was the key Iran hawk in the Democratic Party...
Jordan Schneider: For international listeners, Menendez is now in jail for having taken gold bars from Egypt. But anyways, continue, Eddie.
Edward Fishman: It’s a wrinkle in the story. Then Mark Kirk, who’s his Republican counterpart, who also wants to do a naval quarantine of Iran — the two of them basically say, “We don’t care, Obama, we’re going to sanction Iran’s central bank.” That amendment passes 100 to 0 in the Senate.
Obama is left with figuring out how to make this work. They come to a compromise with the Hill in which they agree to sanction the Central Bank of Iran, but they create two exceptions. One is an exception for countries who every six months significantly reduce their purchases of Iranian oil. For instance, if you’re a Chinese bank, you’re exempt from this — you can pay the Central Bank of Iran so long as China as a whole every six months reduces its overall purchases of oil from Iran. This gives a glide path for Iranian oil sales to decline over time and winds up working marvelously, luckily with the ramping up of shale production in the U.S.
The other exception put in place in 2012 says you can pay the Central Bank of Iran if you’re a Chinese refinery or bank, but those payments have to go into an escrow account that stays inside China and can only be used for bilateral trade between China and Iran.
This actually gives Chinese entities an incentive to comply, because keeping this money in China is going to boost Chinese exports to Iran — there’s nowhere else that the Iranians can use the money.
The one-two punch of these gradual oil reduction sanctions and the escrow accounts leads to a situation where Iran’s oil sales collapse by 60% by volume and it effectively has zero access to its petrodollars. Within 18 months, about $100 billion of Iran’s oil money gets trapped in these overseas escrow accounts. This is the context in which Iran’s economy really goes into free fall. Hassan Rouhani, a dark horse presidential candidate in 2013, won the Iranian presidency on an explicit platform of trying to get the sanctions lifted.
The remarkable thing about this oil sanctions regime is it’s probably the most effective oil embargo we’ve seen in modern history. It’s done unilaterally by the U.S. — no other countries are fully bought into this. It doesn’t involve any sort of naval strategy at all. There’s no quarantining of oil ships or anything. It is just using these threats of being cut off from the dollar to coax banks in places like China and India to comply with American dictates.
Jordan Schneider: This is going to be the poster child for decades of history books in that it actually created political change. It both drove home economically, causing hyperinflation and really hitting growth, and then got you a new slate of politicians who some would argue really wanted to make a deal. Looking back 15 years later, what’s your take on JCPOA and how we should think about the lessons from how the Obama administration used the leverage that they created with this oil embargo?
Edward Fishman: The JCPOA is the high point of American economic warfare in the 21st century in that you actually see sanctions leading to the outcome that the United States had set out, which was to get a peaceful resolution to Iran’s nuclear program. You can quibble about whether the terms of the JCPOA were stringent enough. However, there’s pretty good consensus that sanctions were the critical unlock to that deal.
Democrats say that sanctions were the key to getting the deal. Republicans say that sanctions were working so well that if we had only kept them in place longer, we would have gotten an even better deal. Within really a 10-year period, we flip that consensus from sanctions don’t work to sanctions are this magic bullet that just ended Iran’s nuclear program without firing a shot.
The key lesson here is that you need both economic leverage to make sanctions work and a clear political strategy. Having a clear political strategy, which was to get a nuclear deal with Iran, wound up being very important because you wind up having the international community grudgingly go along with the sanctions. They don’t voluntarily go along — they kind of have to be dragged along, including even the Europeans. But it would have been much harder to bring them along if there hadn’t been a political strategy, if it had just been bludgeoning Iran with economic pain without any sort of political end game in mind.
Responding to Russia (2014 vs. 2022)
Jordan Schneider: Let’s transition from the success of Iran sanctions to the failed response to the annexation of Crimea. What was different about how Obama and the world responded to Russia’s invasion in 2014?
Edward Fishman: Too often we tell our histories in silos — U.S. policy toward Iran vs. U.S. policy toward Russia. One thing I wanted to show in my book is that all of these sanctions campaigns are intertwined because ultimately these are the same decision makers at the table in the Situation Room across multiple issues.
The timeline is interesting here — the U.S. signed the original Iran nuclear deal, which froze Iran’s nuclear program, on November 24th, 2013. On the same exact day, hundreds of thousands of protesters descended upon the Maidan in Ukraine to protest Viktor Yanukovych’s deal with Putin.
The Ukraine crisis really does wind up taking the Obama administration by surprise. It’s not like the Iran nuclear program, which played out over the years as a slow-burning crisis. The Ukraine crisis and the Crimea annexation happened very quickly, with the U.S. constantly playing catch up. This parallel is important because right when Obama officials are scrambling to figure out what to do about Putin’s annexation of Crimea, they’re fresh off this giant victory where they just froze Iran’s nuclear program basically just by using sanctions.
It became natural for Obama officials in February-March of 2014 to say maybe sanctions could work against Russia. It’s a harder problem with Russia for several reasons. Russia has a much larger economy than Iran — in 2014 it was the 8th largest economy in the world and the world’s largest exporter of fossil fuels. Europe is completely dependent on Russian energy to heat their homes. Natural gas pipelines crisscross the continent between Russia and Europe.
Putin is creating facts on the ground as the U.S. is trying to scramble to put together sanctions. The annexation of Crimea happens within weeks of the “little green men” showing up in Crimea — they appear at the end of February and the annexation is formalized in middle of March. Shortly thereafter, Putin starts sending little green men into the Donbas, Ukraine’s industrial heartland.
Jordan Schneider: Let’s focus on the multilateral dynamic of this because obviously the UN is thrown out when Russia’s doing the thing. I remember very vividly watching the transition of the European actors who were pretty close to shrugging off this whole thing — until all those Dutch people died in the commercial liner that the Russians shot down by accident with their anti-aircraft missile. Can you explain how that changed the dynamic?
Edward Fishman: When Putin annexed Crimea in March of 2014, the U.S. and Europe did go ahead with some sanctions, but by and large they’re individual sanctions on people very close to Putin — his judo partners from childhood who have been elevated to positions of power at companies like Rosneft. Igor Sechin, for instance, the CEO of Rosneft, is sanctioned, but there are no sectoral sanctions, no actual significant economic sanctions on the Russian oil industry or its banking sector.
Obama and European leaders very publicly threatened this in March of 2014, but they don’t do anything. The reason is partly because there isn’t political will, but it’s also because they don’t know what kind of sanctions are tolerable to their own economies. They wind up spending months negotiating and coming up with what they eventually term “scalpel-like sanctions,” which effectively cut off big Russian state-owned enterprises from Western capital markets. It’s using an even narrower chokepoint than the dollar — it’s really just Western financing.
Interestingly, something that doesn’t often get recognized enough, the Obama administration went ahead with these sectoral sanctions, cutting off some big Russian energy companies and banks from U.S. capital markets on July 16, 2014, the day before MH17 was shot down. Obama and his team were getting fed up with the European foot-dragging. They say we need to send a powerful signal to Putin if we’re going to have any chance of deterring a broader invasion of the Donbas.
At the time, the New York Times was publishing headlines like, “Obama goes ahead without the Europeans.” Banking CEOs in the U.S. are incredibly upset because they’re saying this is just going to lead to a flight from the dollar to the euro and all our competitors in Frankfurt and London are going to benefit at our expense.
The next day, Putin’s proxies in the Donbas shot down a commercial airliner using a Russian-made Buk missile. They killed almost 300 people, by and large Europeans, most of them Dutch. All of a sudden the political aperture just widens completely in Europe. The Europeans are suddenly not only ready to match the U.S. sectoral sanctions of July 16, but actually go beyond them — they wind up cutting off all of Russia’s state-owned banks from the European financial system. The real core sectoral Russia sanctions are put in place after MH17, really from late July 2014 through September 2014 when Russian and Ukrainian leaders agree to the first Minsk agreement, the first ceasefire in the conflict.
Jordan Schneider: There are two parts that made me get upset rereading and reliving this story. One is that the Obama administration had just learned the lesson which Democrats in general have a really hard time with — escalate to de-escalate. It’s such an Obama thing, the same with the debt ceiling, where he was just like, “I’m going to be a nice normal actor and lay out my five demands and okay, we’ll get to two or three.” The Tea Party — this is ancient history now — and the Republicans were like, “No, we want 100% of what we want.” Obama would get scared, then they’d do a debt ceiling fight and he would end up giving way more than he realized he had to.
By the time we got to 2014, he just said “screw you.” He had the playbook with Iran. All the Treasury forecasting about the catastrophic costs of sanctions is overblown. The U.S. had more agency than expected, the euro was not going to take over.
But Russia really got away without serious economic consequences. Why didn’t Obama put the money where his mouth was?
Edward Fishman: In retrospect, there are two things that led to Obama’s overly cautious approach. One was real, genuine concern about the U.S. economy and the European economy. Remember, we’re still in the wake of the financial crisis and the Eurozone crisis is very much a live situation. There are genuine fears from the Treasury Department that you could accelerate a financial crisis in Europe if Russia were to cut off their gas supplies, and that contagion would spread to the US.
The other thing — this is an interesting paradoxical lesson for the Trump people now and people who say Europe needs to pull more of its own weight — Obama was very deferential to the Europeans over the Ukraine crisis. He explicitly wants people like Angela Merkel and François Hollande to take the lead. The negotiating block that came up with the Minsk agreement, the Normandy format, is France, Germany, Russia, and Ukraine. The U.S. doesn’t even have a seat at the table in the negotiations. Obama was saying, “This is in Europe’s backyard. It’s really their problem.”
In retrospect, that caution does not look very wise. Obama should have hit Russia much harder than he did in 2014. One interesting thing though is even though the sanctions put in place that summer — these capital market restrictions, the “scalpel-like sanctions” — are much weaker than the Iran sanctions, in the second half of 2014, oil prices cratered from over $100 a barrel to around $50 a barrel.
While the sanctions were aimed at trying to constrain Russia’s economic horizons as opposed to creating an immediate financial crisis, the sanctions do push Russia to the brink of a complete meltdown. In the winter of 2014-2015, Russia’s economy looks like it’s about to collapse — honestly just as bad, if not worse than Russia’s economy winds up looking after the much more drastic sanctions from February-March 2022.
The reaction is remarkable. I have some of these quotes in the book. European leaders look at this and say, “This isn’t scalpel-like — this is what we signed up to. We didn’t want to push Russia off a cliff.” Hollande, the French president, actually says, “We explicitly don’t want to push Russia to its knees.” The Europeans, and to a certain extent the United States, got spooked by how impactful the sanctions are because they wind up being accelerated by this collapse of oil prices. Part of the reason why there’s a real frantic desire to get another more permanent agreement, which winds up being called Minsk II in February 2015, is because the Europeans really didn’t want to see Russia’s economy fall off a cliff.
Jordan Schneider: Elections matter and leadership matters. I like that you included so many McCain quotes about the events in both Iran and Ukraine, since he could have been president during these years.
Edward Fishman: One of the key ingredients of the success of Obama’s Iran sanctions is the fact that there’s this bipartisan supermajority in favor of tougher sanctions on Iran. Even if Obama had instincts to be cautious or lawyerly, Congress was passing draconian sanctions laws 100 to 0 over a veto-proof majority. With Russia, you had no sanctions laws at all.
What that speaks to, which becomes more important as our story develops, is that U.S. companies had a lot to lose in Russia. It’s not as much of a political winner for members of Congress and senators to try to layer sanctions onto Russia because they might hurt a company in their state or district. We start seeing that maybe there are domestic political limits to how far the U.S. is willing to go with economic warfare.
Jordan Schneider: Commitment to sanctions is a key factor. Secretary Lew once remarked, “One of the things the Russians would say to me is, ‘We survived Leningrad, we could survive this.’ Their definition of what they were willing to tolerate was well beyond the realm of what we would consider tolerable.”
America’s rich, and the pain that we would end up inflicting on ourselves with sanctions would only be like a half percentage point hit to our quality of life. Whereas Russia is starting from a lower baseline, and sanctions hurt them way more than they hurt us. Yet, we’re not comfortable letting ourselves be pinpricked, even if it’s to save the international order.
You wrote…
“With the loss of the Russian market, Lithuania’s dairy industry teetered on the brink of bankruptcy. When a team of State and Treasury officials met with a Lithuanian dairy farmer outside Vilnius in 2015, they expected her to express frustration. She did, but it wasn’t about her declining business. ‘You should be hitting Russia harder,’ she said.”
It doesn’t come down to economics for a lot of this stuff. There are the political economy games of the Texas senator wanting to help out Exxon or whatever, but it often is a question of moral righteousness. We live in rich countries and we can afford to go without, by and large, way more than that Lithuanian dairy farmer could go without.
Edward Fishman: That’s exactly right, Jordan. One of the macro ironies of the book is, the rise of economic warfare in U.S. foreign policy in the 21st century is partly because military force became politically toxic in the aftermath of Iraq and Afghanistan. As those wars were going south, neither Republicans nor Democrats felt like they could even fight limited military engagements, which is very different from the ’90s when there were all kinds of small wars and U.S. bombing campaigns.
Economic warfare initially is seen as more politically palatable because it’s not hurting Americans — we can sanction Iran out the wazoo and there’s no pain felt at home. But then once you get to Russia and even more powerfully once you get to China, there are real political risks for leaders who impose sanctions on these countries. Even a 10% spike in oil prices or a marginal increase in inflation can become powerful factors in the minds of American presidents and wind up constraining our ability to successfully prosecute economic warfare.
Jordan Schneider: That’s a great point. In the 90s, you had the Taiwan Straits crisis where Clinton threw a carrier there and things calmed down. You had Mogadishu, you had Yugoslavia. But there’s this moment in 2014 where the Ukrainians asked, “Can you give us Javelins, please?” The Europeans said no. Blankets don’t win wars, bullets do.
This is the heartbreaking thing — if Russia believed that the U.S. and NATO were really going to put their money where their mouth was in arming the Ukrainians for war number one, maybe they would have been more concerned — not only about the economic impact, which they clearly underpriced, but also the military impact. We have had hundreds of billions of dollars of armaments go to help Ukraine. It was totally reasonable for Putin, based on the track record of the Obama and Trump administrations, to not expect that to be the response when it came to 2022.
Edward Fishman: Looking at the real error of U.S. policy toward Russia, it’s not necessarily anything that happened in 2014 because we were dealing with a completely novel problem, an unexpected crisis. There was no playbook for sanctions on Russia. This is one area where it’s important to be empathetic to Obama and his top team because it wasn’t easy what they had to deal with. The sanctions they did put in place in 2014 wound up being really impactful — Russia’s economy effectively collapsed that winter.
The bigger indictment on American policy is what happened after February 2015 when the Minsk II agreement was signed. After that, the Obama administration took its foot off the gas on sanctions, basically saying they’re just going to maintain what they have in place. Russia very publicly interferes in the 2016 election. Obama had threatened Putin with drastic sanctions if he continued to interfere. Putin continued to interfere, and the sanctions Obama put in place in December right before he left office were really minor. That’s a bad signal.
Then you have four years of the Trump administration in which Trump does nothing on Russia sanctions. It’s a logical lesson for Putin to draw, both from the last year and a half of Obama and all four years of Trump, that he basically got away with the annexation of Crimea at a reasonable cost. That’s just speaking of the U.S. — Europe is even worse. In 2015, after the annexation of Crimea, a consortium of companies signed the Nord Stream 2 pipeline deal to double the amount of gas that Europe would get from Russia. Putin was completely within reason to assess that the West does not have the stomach for a real economic war.
Jordan Schneider: Unlike in Crimea, the U.S. sees this coming in 2022 and has months to try to get its ducks in order, to try to do everything it can to dissuade Putin from trying to take Kyiv. What happened then?
Edward Fishman: When Biden comes in, there’s a real debate amongst his advisors about what to do. Russia had accumulated all of these misdeeds that had gone unanswered. Biden himself, when he was vice president, wanted to arm the Ukrainians. He was the most hawkish member of the top Obama team on Russia, always in favor of tougher military steps to help the Ukrainians, always in favor of tougher sanctions.
There was real debate about what to do. Should they come in right away with really tough sanctions? Biden’s conclusion was that we were still reeling from the COVID pandemic, we had climate change to deal with, and China was the biggest geopolitical issue on his radar. They tried to have what they called a “stable and predictable relationship” with Russia — which is hilarious in retrospect, as “stable” and “predictable” aren’t things you necessarily ever ascribe to Putin’s Russia.
They came out of the gate in April 2021 with a modest increase of sanctions, saying, “Here’s some sanctions to repay you for all these bad things you’ve done over the last six years. But after this, we want stability and predictability.” Putin gets a summit with Biden, which he’s very happy to get. Then he pens a rambling 5,000-word essay about why Ukraine’s not a real country and should be part of Russia in the summer of 2021 while he’s in lockdown. He masses over 100,000 troops around Ukraine’s border that fall.
It becomes quite clear that Putin has designs on Ukraine. In what is probably the biggest intelligence success of the 21st century, the US intelligence community gets Putin dead to rights. They figure out exactly what his plan is, to the point where Biden starts warning American allies privately in September and October 2021 that an invasion is coming. Very soon thereafter, he starts making public warnings that invasion is coming and tries to use the threat of swift and severe consequences, particularly very dramatic economic sanctions, to deter Putin from invading Ukraine.
Jordan Schneider: Let’s talk about how they tried to build that coalition and signal those sanctions in the lead-up to the ultimate invasion.
Edward Fishman: A stroke of luck for the Biden administration was having Daleep Singh, who had played a significant role in the 2014 sanctions. He’s one of the top financial minds in Washington — a city that doesn’t have many people with deep financial markets expertise. Daleep is an exception. He was in the perfect role to orchestrate a sanctions campaign as the Deputy National Security Advisor for International Economics, overseeing the organs of the US Government that do economic warfare.
In late 2021 and early 2022, Daleep builds relationships with his fellow G7 counterparts: in Brussels, Bjoern Seibert, and in London, Jonathan Black. They start getting into the nitty-gritty of what kind of sanctions they might impose if Putin were to invade. This preparation is important not just for being ready to do something real if Putin pulls the trigger, but also for making the threat of deterrence more credible. Russia has a world-class intelligence apparatus — if all you had was Biden wagging his finger saying “You’re going to face really strong sanctions if you invade,” but there’s no actual bureaucratic movement in these capitals creating sanctions ready to go, Putin would probably assess it was a bluff. The preparation that Daleep Singh and his counterparts in Europe and Japan do is very important.
Jordan Schneider: I love how they were doing this like in secret, but also in public. They weren’t being super hard about using classified communications — they were just calling each other on their phones because they actually want the Russians to be listening and believe they are going to put real sanctions on them.
Edward Fishman: That’s exactly right. They view the preparations as important from both a practical standpoint and a signaling standpoint.
By the time we get to the moment of decision in late February, it becomes clear after Putin and Xi Jinping meet in early February that an invasion probably won’t happen until the Beijing Olympics wraps up — Putin doesn’t want to spoil Xi Jinping’s party. By that time, you have a very extensive menu of sanctions options. Most importantly, you have what’s called the Day Zero package — the raft of sanctions that would go into effect as soon as Putin invades.
The compromise is made because inflation is at a four-decade high and there are concerns about oil prices potentially spiking. Biden says they’re going to maximize sanctions on Russia but not aggressively target its oil sales, which is tough because Russia’s economy depends on hydrocarbon exports. The strategy of the Day Zero sanctions is to implement maximalist sanctions on Russian banks — Sberbank and VTB, the two biggest banks in Russia — as well as Russia’s access to foreign technologies. They took the Foreign Direct Product Rule that had been imposed on Huawei in 2020 and recast it to cover the entire Russian economy. They take something that had been previously employed on just one Chinese company and apply it against an entire state.
The tragedy of the situation is that Putin invades and very quickly — similar to that moment in July 2014 after MH17 was shot down — there’s a giant shift of the Overton window in Europe. Everyone becomes gung-ho for very aggressive sanctions after Putin invades and we start seeing just how horrible this war is and how imperialistic Putin’s goals are. Hundreds of thousands of people protest on the streets of places like Berlin, and there’s a massive political movement in favor of stronger sanctions.
Within 24 hours of the invasion beginning, the Day Zero package that Daleep Singh and his colleagues had worked months on looked much too weak and actually undershot the political moment. Within that first weekend of the war, the United States and the G7 agreed to go much further and actually sanction Russia’s central bank directly — something that was seen as too politically radical to even consider in the lead-up to the invasion. Putin clearly agreed because he had left half of his central bank reserves completely exposed to Western sanctions.
Jordan Schneider: This goes back to the mafia diplomacy concept. Ironically, Putin expected the West to be more gentlemanly and concerned about the centrality of the dollar and euro to global trading. Once the war started and the Overton window shifted — which everyone had a hard time foreseeing — things changed. Looking back, it seems silly that they didn’t anticipate massacres when Russia invaded. While sanctioning their central bank was an option, there remained questions about whether they could get the money out, and if they would even believe the threat before it happened. The actual deterrent value we had during those months remains an open question.
Edward Fishman: Clearly, we would have been better off had the U.S. and Europe created more aggressive sanctions plans in advance. This could have strengthened deterrence and weakened Russia’s economy and warfighting capability more quickly, directly helping Ukraine on the battlefield. There were significant costs to underestimating how willing political leaders would be to implement tough sanctions in the U.S. and Europe. But going back to your earlier point, Jordan — from a deterrent standpoint, would that preparation have overridden Putin’s lesson from 2014 and the seven or eight years of basically allowing Russia to get off scot-free after annexing Crimea? Putin had likely already sized this up in his head by then, and I’m not sure we could have changed his mind.
Jordan Schneider: Here’s a crank idea — why didn’t the Treasury Department go long on oil if they were worried about it spiking up to $250 a barrel? Couldn’t you just do the math that way?
Edward Fishman: This is a point I make toward the end of the book — the U.S. is much better at imposing economic penalties than deploying capital for strategic reasons. That would be a very creative use of government resources, but it’s not a bad idea. If we had the flexibility to do something like that in a strategic manner, sure. We do use things like the Strategic Petroleum Reserve to stabilize the oil market. In March 2022, the Biden administration released 180 million barrels of oil to try to stabilize the market.
Jordan Schneider: They did eventually act, but it took too long, and the Department of Energy people are complaining that the caves might crater in. Reading through your book, I can only imagine how frustrating it must be for these officials working around the clock to get the whole world to ramp up sanctions, and they can’t even get their own government to release oil for arguably the biggest crisis in at least 50 years.
Edward Fishman: Many of our institutions are built on the assumption that we live in a peaceful, predictable world, and we don’t always get our act together in time for crisis. This isn’t unique to the 21st century — it’s been true throughout American history.
Jordan Schneider: Here’s another crank idea for you. In the winter of 2023, everyone was terrified that oil prices were going to spike. Did anyone discuss geoengineering solutions, like spraying sulfur in the air over Europe to save everyone’s energy bills?
Edward Fishman: There are a number of tragedies in this story, one being that you decided to become a podcaster instead of a sanctions nerd. Had you gone down this path, maybe we would have benefited from your creativity in the U.S. government.
Institutional Dysfunction
Jordan Schneider: The people you profile, whom you clearly admire for their incredible feats of civil service, were creating new concepts and regimes unimaginable back in 2004 while operating under such constraints in such a dysfunctional system. They made enormous family sacrifices, which you mention several times. We did a show called “Is the NSC Unwell?” where we opened with Jake Sullivan being awake at 4 AM on a Tuesday during a home invasion because he was dealing with Ukraine issues.
Having the idea is the easiest part. Sure, I can suggest geoengineering to fight the impact of Russian oil, but transforming a clever idea that checks all the economic, institutional, and diplomatic boxes into reality is unbelievably difficult. Multiple times in your stories, there are eight-month delays for things that everyone should have immediately approved on day one.
Edward Fishman: We need a government that’s purpose-built for the age of economic warfare. That’s the premise of my book — we are living in an age of economic warfare. Sanctions, tariffs, and export controls are how great powers compete today and will compete tomorrow. This is a secular trend we’ve seen throughout the 21st century, yet we haven’t changed our government to actually fight and win these economic wars.
There’s nothing like the Pentagon for economic warfare. During my short stint at the Pentagon working for then-Chairman of the Joint Chiefs of Staff Marty Dempsey, I noticed that military force has one agency and a clear chain of command up to the Secretary of Defense. With economic power, you’ve got numerous agencies involved — the Treasury Department, the Commerce Department, the State Department, the Energy Department. Much time is spent just coordinating the interagency process.
Ideally, we would have a dedicated department with clear leadership for economic statecraft or economic warfare. Some governments have moved in this direction — Japan now has a cabinet-level minister for economic security. The U.S. hasn’t innovated like that. There’s a core budgetary problem where agencies like TFI (Office of Terrorism and Financial Intelligence) at Treasury, which Stuart Levey led, or BIS at the Commerce Department, haven’t seen significant budget increases despite their missions growing exponentially.
Jordan Schneider: This theme comes up repeatedly in these stories and with the chip export controls. When cabinet-level officials disagree without presidential direction saying “We’re doing X, not Y, get with the program,” things stall or take longer. Cabinet members are congressionally approved; their words carry weight. When Janet Yellen believes a sanction would harm global inflation and the American economy, Jake Sullivan must call Mario Draghi to persuade her because Biden won’t act without her support. Everyone has different priorities, and without a central authority or an engaged president, you end up with stasis — allowing Russia to make an extra $200 billion they shouldn’t have throughout 2023.
Edward Fishman: Exactly. The Draghi call is one of the more remarkable episodes in the book. After the political aperture expanded during the first weekend of the Ukraine invasion in 2022, making central bank sanctions possible, the G7 agreed. Then Janet Yellen raised concerns, requiring a call from Mario Draghi, Italy’s leader and former European Central Bank chair, to personally assure her it was acceptable.
Regarding China, much of why your podcast is amazing has been its in-depth coverage of chip export controls. Looking back to the first Trump administration, export controls were deployed against Huawei instead of sanctions largely because Treasury Secretary Steven Mnuchin opposed a tough China policy. In early 2019, after the arrest of Meng Wanzhou 孟晚舟, some administration officials suggested sanctioning Huawei and putting them on the SDN list. Mnuchin refused, so they defaulted to putting Huawei on the entity list, which Wilbur Ross controlled as Commerce Secretary. The whole export controls landscape might have been very different with a more hawkish Treasury Secretary during the first Trump administration.
Jordan Schneider: You have this wild anecdote from Matt Pottinger, former ChinaTalk guest who became Deputy National Security Advisor towards the end of the Trump administration.
Pottinger noted that at one point, Bolton decided not to tell Trump about arresting Meng Wanzhou. Pottinger interpreted Trump’s rhetoric as supporting a tough stance on China.
“Pottinger told his Commerce colleagues that Trump was pursuing a two-pronged strategy. On the one hand, the president was seeking to preserve his personal relationship with Xi Jinping and the appearance of pursuing warmer ties. But as for officials in the bureaucracy, Trump ‘wants us punching as hard as we can.’ In effect, Pottinger was telling the Commerce officials to take Trump seriously, not literally — to tune out the verbal concessions that Trump made in public and keep a default position of being ‘tough’ on China.”
Presidents, even those not in their 70s, only have maybe 5% of their day for these matters. This leaves an enormous amount to be sorted out by empowered appointees and cabinet members, which explains how we ended up with export controls instead of sanctions on Huawei — quite remarkable in retrospect.
Edward Fishman: The first Trump administration has been characterized as super hawkish on China, but examining the record shows Trump himself wavered between being very hawkish and totally obsequious to Xi Jinping. The policy was shaped by different factions: people like Pottinger and Bob Lighthizer were tough on China, while Mnuchin and Gary Cohn wanted to return to the early 2000s approach — the Hank Paulson school of U.S.-China relations. These factions took advantage of opportunities when Trump leaned their way to advance their policies. Trump didn’t take a more consistently hawkish line toward China until his final year in office, when he believed Xi Jinping had lied to him about COVID, destroying his re-election chances. We’ll likely see similar dynamics in a new Trump administration — Trump vacillating while different factions capitalize on moments when he’s more receptive to their proposals.
Jordan Schneider: You close the book, Eddie, with the idea of an impossible trinity.
“We don’t yet know when the Age of Economic Warfare will end, but we can envision how. The trade-offs facing policymakers in Washington, Beijing, Brussels, and Moscow can be thought of as an impossible trinity consisting of economic interdependence, economic security, and geopolitical competition. Any two of these can coexist but not all three.”
Walk me through the 20th and 21st centuries — what different trade-offs did states make, and where are we landing now in 2025?
Edward Fishman: Let me explain why I ended the book this way. While I wrote a narrative history because I believe individuals can shape history — remove certain individuals and history would have gone differently — there are also structural reasons underlying the age of economic warfare. Consider this statistic: Barack Obama used sanctions about twice as much as George W. Bush, Trump used them twice as much as Obama, and Biden uses them twice as much as Trump. This suggests both individual agency and structural factors matter.
The geoeconomic impossible trinity I developed explains why this is happening. You can only have two of these three elements simultaneously — economic security, economic interdependence, and geopolitical competition. During the Cold War, we had economic security and geopolitical competition in a bipolar order between the U.S. and Soviet Union, but at the expense of economic interdependence — there was no meaningful economic relationship between them.
When the Cold War ended, geopolitical competition disappeared. China and Russia transformed from adversaries to potential friends, and we invested significant political capital bringing both into the liberal international order, including the WTO and other key international bodies. Without geopolitical competition, we could embrace economic interdependence without sacrificing economic security.
Today, we maintain economic interdependence while geopolitical competition has returned full force, resulting in lost economic security. This affects all major powers — the United States, Japan, European Union, China, and Russia. None feel economically secure, leading them to invest heavily in protecting themselves from rivals’ sanctions, export controls, and tariffs. To regain economic security, we must either end geopolitical competition, which seems unlikely, or significantly reduce economic interdependence. My view is we’re heading toward a significantly less interdependent global economy in the years ahead.
Jordan Schneider: You end the book with some dark words,
“Without the ability to channel geopolitical conflict into the economic arena, great powers could once again find themselves fighting on an actual battlefield. The dream of economic war, for all its downsides, is that it can be an alternative to a more violent kind of war. Someday the age of economic warfare might end, but we might miss it when it’s gone.”
Care to elaborate on this idea?
Edward Fishman: We face very significant stakes in our economic decisions today as we head toward a less interdependent global economy. This could manifest in two ways. First, a world economy where the U.S. and its allies deepen their connections. We might have less trade with China and Russia, but more with Canada, Mexico, the European Union, and Japan. Janet Yellen in the Biden administration called this “friendshoring.” Bob Lighthizer proposed this in a recent New York Times op-ed, suggesting the U.S. and other democracies create a bloc with low internal tariffs and high tariffs on everyone else.
The alternative is deploying sanctions, tariffs, and export controls arbitrarily against friends and foes alike, creating a chaotic breakdown of the global economy. We’d be forced into autarky by default, without long-term economic agreements with allies or adversaries. This scenario frightens me most because history shows that when states can’t secure resources and markets through free trade and investment, the temptation for conquest and imperialism rises.
President Trump’s talk about seizing Greenland for its mineral resources echoes Hitler’s pursuit of Lebensraum. Hitler feared being cut off from European trade after Europeans sanctioned Mussolini for seizing Abyssinia. If economic interdependence unravels into every country for itself rather than friendly blocs, we could see a return to great power war.
Jordan Schneider: Dark. I’ll refer folks back to our two-part episode with Nicholas Mulder on The Economic Weapon, which told that whole 1920s and 1930s story of how Imperial Japan and Nazi Germany developed their autarkic, resource-hungry vision. While racial ideology played a role, they were clearly terrified about accessing enough oil, minerals, and resources to remain great powers.
Researching Modern History
Jordan Schneider: Let’s shift topics. Tell me about writing history of the past 20 years. You don’t have everything declassified, you’re doing interviews, and history seems to be happening in WhatsApp groups. What was it like both as a former civil servant and then interviewing all these people to piece this recent history together?
Edward Fishman: As you know, Jordan, since we shared some classes, I studied history and in a parallel universe might be a university historian. After college, I went into government work and realized that in this era, many decisions bypass formal processes. Even back in the 2010s, decisions were made through informal communications, in coffee shops, never written down, through WhatsApp groups. This has only accelerated since I left government.
Contemporary history plays a crucial role because documentary records won’t be as valuable in 30 years as they were previously. They might even mislead — often the package going into an NSC meeting doesn’t reflect what’s actually discussed or decided. Many decisions happen outside formal meetings entirely.
This experience convinced me that the best approach was to follow Thucydides’ method — write contemporary history, documenting the times you live in, striving for impartiality. What you lose in documentary records, you gain by talking to people who were actually present. Thanks to my government experience and non-partisan reputation, I accessed everyone crucial to this story — Democrats, Republicans, and current civil servants.
Future historians will surely build on and improve the story told in Chokepoints when they access all documents. However, I hope the insights derived from my access to these people and my insider government experience will prove durable.
Jordan Schneider: Did you send Nabiullina an email?
Edward Fishman: No, I didn’t speak to Elvira Nabiullina, unfortunately. One wrinkle in the story is that I was sanctioned by the Russian government in 2022, before I even started writing. I’m currently banned from any travel to Russia.
Jordan Schneider: She’s got an open invitation to ChinaTalk. I’d love to hear her side of the story.
y through declassified documents showing what really happened — I’d bet most of the narrative around U.S. policy holds up. Rather, I hope we’ll see Chinese, Russian, or European versions of Chokepoints. While I capture those stories to some extent, the book focuses on the United States. If counterparts in those systems wrote similar books, we’d have a much more complete picture.
Jordan Schneider: Eddie and I were classmates at Yale, studying ancient history together. I love how you say you’re walking in Thucydides’ footsteps — let’s say we’re doing the same with ChinaTalk. For both of us, Donald Kagan’s classes were among the most formative in thinking rigorously about politics, history, and warfare. Any memories or reflections about his impact in the classroom?
Edward Fishman: One sad aspect of publishing this book is that Don died a couple years ago and won’t have the chance to read it. Of all my teachers, he had the biggest impact, shaping my career in many ways. He even influenced how I teach my class at Columbia on Economic and Financial Statecraft — I use his exact seminar format, with students debating each other’s papers weekly.
The main lessons I learned from Kagan that influenced the book include understanding the role of contingency in history — people and their decisions matter. While many history books focus on impersonal forces, Kagan taught me that structure sets context but free will and decisions can change history’s course. That’s why I focused on the people creating these policies.
Second, chronology matters. You must understand historical decisions within the knowledge available at the time. We tend to judge past decisions with hindsight, but understanding what people knew then reveals more about how history unfolds.
Finally, history itself matters. Kagan said, “Without history, we are the prisoners of the accident of where and when we were born.” Beyond clichés about repeating history, understanding what our predecessors did right and wrong helps us live better lives today.
Jordan Schneider: Another lesson coming through your book is that while we can debate grand strategic decisions, like Biden’s approach, the most human agency appears one or two levels below. Having someone from Goldman Sachs who understands the global insurance market enables implementing policies that might not otherwise be conceived. While we criticize civil servants in today’s America, it’s important to recognize that you can expand government’s effectiveness by empowering the right people to make decisions and analyze questions thoughtfully. For anyone at a career crossroads, read Eddie’s book and understand that your future choices matter.
Edward Fishman: I appreciate that, Jordan. If there’s one takeaway, it’s that government officials’ decisions truly matter. The protagonists I highlighted — Stuart Levey, Adam Szubin, Dan Fried, Matt Pottinger, Daleep Singh, Victoria Nuland — if you remove them from their situations, you’d have very different policies. We were fortunate to have them in those positions. Having more people with diverse skill sets willing to serve in government increases the odds of having the right person in the right place at the right time.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
Mood Music—Iranian Pre-Revolution Psychadelic Rock
Login
