Jordan Schneider: Tony, give us the 101 on what a National Security Strategy is, and then we’re all going to go around and say one nice thing about it.
Tony Stark: There are three major U.S. government national security strategy documents. The first is the National Military Strategy, which applies to the uniformed services but is rarely noticed outside the Joint Staff.
Next is the National Defense Strategy (NDS), which is the Pentagon’s primary strategic document. It’s the one most people in the field care about because it’s a Cabinet-level document, even if it isn’t overtly political. Legally, a new NDS is required every four years, and developing a new NDS takes 6 to 18 months. New administrations are given a little extra time — about a year and a half — to publish their first one.
The NDS is written at the “action officer” level, which includes General Schedule (GS) employees, field-grade officers, contractors, and think tank experts. Then it is passed up to the Deputy Assistant Secretary level in the Office of the Secretary of Defense (OSD) — their equivalents are three-star generals — and then to the commands, the undersecretaries, and so on.
Finally, there’s the National Security Strategy (NSS), which is historically the most political of the documents because it comes out of the White House, not the Pentagon. The NSS is a guiding vision of the administration’s goals and incorporates all elements of national power. Historically, this is also the blandest document — its wide scope reads more as a political statement than a defense plan. The new Trump administration just released its first NSS. While the NDS has been ready for a while, they were likely waiting to publish the NSS first.
At 29 pages, the new NSS is the right length for a public national strategy document. There are usually non-public, classified annexes and other materials.
Justin McIntosh: The document correctly focuses on economic re-industrialization and re-energizing the defense industrial base — issues we’ve previously discussed. It puts those ideas forward in its “answers” section. But…
Jordan Schneider: No “buts.”
Justin McIntosh: Okay! Yes, that’s where the focus should be.
Jordan Schneider: The straightforward questions in the document are nice. The Q&A rhythm is interesting and provocative. It’s focused. There’s a section of questions like, “What should the U.S. want overall?” and “What does the U.S. want from the world?” There’s no artifice about how transactional it’s going to be — what you see is what you get.
Tony Stark: If I were framing a strategy document for the American people, this is how I would structure it. A clear layout saying, “This is what we want. This is why we have a strategy. What are the ends, ways, and means? What does that mean?” It’s written in a clear, accessible way, without many buzzwords. Although what replaced the buzzwords wasn’t great.
Jordan Schneider: Avoiding policy jargon in this document seems to have been a conscious choice.
Justin McIntosh: But it lacks nuanced, impartial language and contains statements that our adversaries will exploit. A comment on the necessity of securing borders said that any sovereign nation has the right to control them. The PRC and Russia can easily seize on a statement like that. This is a kind of language previous administrations have avoided, because they didn’t want a quote interpreted as agreeing with the Chinese or Russian position.
Trump and Xi chat in South Korea, October 30, 2025. Source.
Tony Stark: The document does not change U.S. policy towards Taiwan. If anyone tells you it does, they are wrong. However, it does give the PRC political and legal ammunition. They can now say, “But you said you wouldn’t interfere in the internal affairs of others,” pointing to our supposed principles of non-interventionism.
The document also says we do have to intervene sometimes. This amounts to talking out of both sides of your mouth — we reserve the right to do whatever we want. The “flexible realism” section is a fancy way of saying we’ll do whatever is convenient. Historically, that has been U.S. foreign policy in practice, but that doesn’t mean it’s what we should aspire to.
Justin McIntosh: I don’t have a problem with them laying out the “ends, ways, and means” discussion up front, but it has limitations. That linear framework is well-suited to military decision-making, but a national strategy needs to be more pragmatic and flexible. At the national level, you control all the resources. You can marshal all those resources toward any goal that is deemed important. That makes the “ends, ways, and means” calculation irrelevant because you will find a way to make it happen.
Jordan Schneider: The Trump administration’s focus on “ends, ways, and means” raises the question — how weak do they think the U.S. really is?
Reducing the U.S.’s power to an “ends, ways, and means” calculation only works in military contexts — counting ships and battalions to see how many wars you can fight. The U.S.’s power to achieve economic and national security ends is elastic. The means to those ends can grow dramatically when the president builds a consensus around them — once the nation decides something must be done, it finds the capacity to do it.
It’s a mistake to define goals downward because those goals inevitably change. Consider the border — the Biden administration didn’t prioritize the issue and struggled to find the means. The Trump administration’s intense focus on the border unlocked congressional funding and operational capacity. The resources didn’t appear from nowhere — the will to use them did. This dynamic applies globally. To believe the U.S. cannot act because it lacks on-hand capabilities is a severely limited way of thinking about our power to shape events.
Mixed Signals
Tony Stark: The document’s focus on military and economic power isn’t unique, but its goals do not align with a realistic budget. It calls for both bolstering deterrence in the Indo-Pacific and shifting our entire global military posture to the Western Pacific, which would drain resources from Europe and Latin America. We have to assume this will happen.
This creates deep concern for our allies, but that matters for the U.S. too. The Germans will be wildly pissed about how they are described in the document. Asian allies are told to “do more,” a demand that ignores their significant recent efforts. Getting allies to increase defense contributions was an accomplishment of the first Trump administration that continued under Biden. The call to “do more” is now an outdated talking point — they are doing more. Japan is considering exporting weapons for the first time.
Justin McIntosh: Worse still, when allies make the kinds of statements the U.S. wants — like Sanae Takaichi declaring a PLA incursion into Taiwan a national security threat to Japan — the administration’s response is silence. Based on the reporting of Xi and Trump’s call, it appears the U.S. did not affirm that position. Instead of backing Japan’s strong stance, the message was to “calm it down.”
The Trump administration is sending mixed signals. Does it want allies to spend more on defense, develop a stronger defense mindset, and care more about their own security, or not?
Jordan Schneider: Let’s do some reading from the scripture here.
“A favorable conventional military balance remains an essential component of strategic competition. There is rightly much focus on Taiwan, partly because of Taiwan’s dominance of semiconductor production, but mostly because Taiwan provides direct access to the second island chain and splits Northeast and Southeast Asia into two distinct theaters. Hence, preventing a conflict over Taiwan, ideally by preserving military overmatch, is a priority. We will also maintain our long-standing declaratory policy on Taiwan, meaning that the United States did not support any unilateral changes to the status quo in the Taiwan Strait.”
From that, it sounds like a good idea for Japan to make its role in deterrence transparent. How seriously should we take any of these documents?
Tony Stark: I wish Eric were here for another briefcase-carrier rant. In the 2010s, a gripe of mine was hearing mainstream national security people, the ones in the know, say strategy documents don’t matter. That is a clear indicator they either haven’t written a good strategy document or haven’t marshalled the resources and people to execute it. I’ve occasionally had to metaphorically beat somebody over the head with a strategy document.
One problem is that people don’t read strategy documents. I have been in meetings with theater-level commands who’ve asked me, “What are you quoting from?” And my response is, “The National Defense Strategy.” They’ll ask me to send it to them. It’s a public document.
Justin McIntosh: “No, no, we meant the classified annex, Tony. Obviously, we’ve read the public one.”
Tony Stark: “The super-secret one that wasn’t even fully distributed to your command.”
Justin McIntosh: The document doesn’t matter, and there isn’t a robust national security apparatus anymore — at least in this administration — it’s as if the President is the sole decision-maker. Trump has consolidated his counsel — it’s a smaller group than it was.
Another problem is that the strategy document’s promises are often the opposite of what the president himself has done. The strategy specifically addresses deterring propaganda aimed at Americans, clearly referencing China, and yet TikTok is still legal here.
When X turned on a filter showing where accounts came from, it revealed so-called Mongolian accounts weren’t Mongolian, and supposed Uyghur accounts were run from mainland China. Pro-MAGA accounts were operated from VPNs in India and China to target Americans. Where was the action on that propaganda? We kept TikTok, and no one has suggested the government force X to shut down foreign influence accounts. These goals are in the document, but the follow-through is missing.
Tony Stark: Every administration struggles with inconsistencies between its strategy and actions. That’s the nature of a democracy — it’s the nature of any government worldwide. This strategy document’s main issue is its unusual use of national security language. The strategy says the administration opposes disinformation, but what do they consider disinformation? There are direct quotes that frame concepts like “de-radicalization” and “protecting our democracy” as a fake guise — that inclusion is wild.
On foreign policy, the document critiques the U.S. for focusing too much on projecting “liberal ideology” into Africa — it’s unclear if that means big ‘L’ or small ‘l’ liberal. Let’s assume it’s both. The most stunning part is that the National Security Strategy of the United States explicitly frames the concept of “protecting our democracy” as a ruse. That is insane.
The parts of a strategy document that truly matter are the ones that diverge from the previous strategies. While I’ve critiqued previous strategies, this document is on another level.
Justin McIntosh: The large section on China is a good example. It would be great if the administration enacted many of the listed actions — I’d be all for it. The cognitive dissonance between the strategy document and the administration’s actions is troubling.
Jordan Schneider: Six months ago, the AI action plan included interesting language about new export controls on semiconductor manufacturing equipment. Those controls are paused because Stephen Miller’s current job is to avoid upsetting China. This directive came after a Chinese official was angered by a Financial Times article on Alibaba and the PLA. Stephen Miller’s Twitter banner is a picture of him shaking hands with Xi. This is hard to square with official strategy documents demanding military overmatch.
You can try to connect those dots and argue that the goal is to keep the economic relationship calm while we re-industrialize and build up our military. Okay, maybe. But that still doesn’t explain the U.S NSS includes sovereignty language seemingly copied and pasted from Putin’s playbook.
Traditional Values, Universal Wings
Tony Stark: The document is also very undergraduate. That is not a critique of the accessible language — I also try to write for a wider audience — but of the concepts themselves. If an undergraduate at the University of Texas at Austin were assigned the paper topic — what should a national security strategy be — this would be that paper.
Jordan Schneider: There are 14 bullet points where each sentence is about seven words long.
Tony Stark: What does this all mean? The language in the National Security Strategy should not shock anyone — it’s consistent with the administration’s usual rhetoric. What has changed is that this language is now the official guidance — it has leverage in bureaucratic fights. The influence may not be immediate, but it will be cumulative. The real test will be when the National Defense Strategy comes out. Someone who worked on it texted me last night and said, “Well, they set the bar low, so this is great for us.”
Justin McIntosh: They’re being pragmatic. What troubled me was the traditionalist language at the end.
“Finally, we want the restoration and reinvigoration of American spiritual and cultural health, without which long-term security is impossible. We want an America that cherishes its past glories and its heroes, and that looks forward to a new golden age. We want a people who are proud, happy, and optimistic that they will leave their country to the next generation better than they found it. We want a gainfully employed citizenry—with no one sitting on the sidelines—who take satisfaction from knowing that their work is essential to the prosperity of our nation and to the well-being of individuals and families. This cannot be accomplished without growing numbers of strong, traditional families that raise healthy children.”
Tony Stark: “We will use every means to protect our precious bodily fluids.”
Jordan Schneider: Wait, if you’re raising a disabled child, or if your child is sick with a fever, then you are not contributing to the restoration of American cultural and spiritual health? Wow.
Tony Stark: That is what RFK Jr. said — if your kid is sick, that’s not a good societal contribution.
Justin McIntosh: His miasmas are off, or whatever non-germ-theory medicine he peddles but doesn’t practice.
Tony Stark: The Midi-chlorians from Star Wars.
Justin McIntosh: That language is reminiscent of what you see from Putin and China’s family planning policies. It is the exact type of language that Xi and Putin use to justify pro-natalist policies and promote traditional families and traditional gender roles. Reading about the one-child policy in Dan Wang’s Breakneck is heartbreaking if you have children. It’s striking how similar the NSS’s language is to China’s early discussion of the one-child policy.
Tony Stark: In a reasonable time, there would be ten articles asking, “What does this mean? How is the government going to encourage people to have more kids?” Now, it’s something I don’t even want to read about.
After COVID-19, as the “China Rising” narrative was gaining prominence in 2021 and 2022, discussions began in national security circles about how the U.S. population is numerically outmatched. Although we are solving that problem with robotics, it was a talking point among traditionalists. They argued that the U.S. won the Cold War by embracing traditional values. That’s not how we won. We won thanks to Skunk Works and the Soviet Union’s economic mismanagement.
This argument has surfaced before in national security circles — it’s not a new phenomenon. The other common concern is protecting our food supply — I’m surprised it was not mentioned in the document. But, to quote a former coworker of mine, “We have Buffalo Wild Wings and the Chinese don’t. I think we’re okay.”
Cheerleaders perform during a baseball game at Taoyuan International Baseball Stadium in Taiwan, May 2018. Source.
Jordan Schneider: That would be a great cultural export. Maybe that’s what the world needs.
Tony Stark: Are there Buffalo Wild Wings locations in Shanghai or Beijing?
Justin McIntosh: I’m sure there’s one in Taipei. [Note from Lily: Taiwan does not have a Buffalo Wild Wings, but it does have three Hooters locations.]
Tony Stark: Is the food different, or is it universal?
Justin McIntosh: It’s universal, but like McDonald’s in Japan, it’s better.
Tony Stark: Another American cultural victory. We don’t need to change anything.
Justin McIntosh: You can watch a baseball game while eating Buffalo Wild Wings in downtown Taipei.
Tony Stark: During COVID, my former American University professor, Justin Jacobs, uploaded all his lectures on Spotify — excellent lectures on the history of China and Japan. He has an episode about why baseball is played in Taiwan but not on the mainland. He discusses the Japanese occupation of Taiwan and the differences in Confucian culture and masculinity. Prof. Jacobs is an amazing resource for East Asian history.
Jordan Schneider: I asked Gemini what other regimes this resembles. It suggested Vichy France, Fascist Italy, and modern Hungary.
Justin McIntosh: I wonder what Grok would say…
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
This December, we’re teaming up with GiveDirectly to send cash to 800 impoverished families in the Bikara region of Rwanda. Studies show that direct cash transfers have a multiplier effect of 2.5x in local economies and reduce infant mortality rates by 48%. Your donation is also tax-deductible in the United States. The link to give is here, and the deadline for donations is midnight on December 31st. Please consider donating if you can!
is a researcher focused on semiconductors, AI, China, and Taiwan. He holds a Master’s degree in Regional Studies — East Asia from Harvard and was recently a summer fellow at the Centre for the Governance of AI (GovAI).
High-bandwidth memory, or HBM, remains the key bottleneck for China to catch up in manufacturing advanced AI chips. As Moore’s Law has more or less held steady, logic nodes have continuously progressed.
However, the rate of memory chip progression has been slow compared to logic chips. Thus, AI operations are often “memory constrained,” meaning that compute is sitting idle waiting for the memory chip to feed it data on which to perform operations. HBM was created to address this “memory wall” by stacking multiple memory chips on top of each other to boost memory bandwidth. As AI chips continue to get better, HBM remains a critical component for scaling. Simply put, if you care about the AI race and AI chips, then you must care about HBM.
Although China’s memory champion CXMT has been closing the HBM gap, the three memory giants of SK Hynix, Samsung, and Micron continue to be more than two generations ahead of CXMT’s HBM2. Assuming export controls hold steady, China’s HBM advances will continue to be stymied by a lack of advanced equipment.
For perspective, achieving the industry’s current HBM3E and HBM4 would be a tremendous achievement for China. As of November 2025, the most advanced AI chips in use use HBM3E. H100s, B100s, and other leading GPUs tap into HBM3E for memory, while Nvidia’s upcoming Rubin GPUs will use HBM4. If CXMT can achieve HBM4 quickly, then they will be able to crack a key part of making advanced GPUs. However, even if they are able to make HBM4 several years down the line, competitive AI chips will likely have meteored beyond contemporary standards to handle workloads unimaginable today.
Ray Wang’s piece earlier this year in ChinaTalk mapping CXMT alongside other memory giants helps policymakers keep an eye on China in the rearview mirror. But past HBM2, when will CXMT hit a wall? Given the current state of export controls and Chinese technological development, what node of HBM can China be expected to reach?
The Three Ingredients: DRAM, Base Die, and Packaging
Making HBM is a difficult endeavor, and the product’s performance ultimately comes down to three factors: the DRAM dies that compose the HBM, the base die that routes the signals coming in and out of the memory stack, and the packaging that binds the DRAM dies together.
Different bottlenecks exist within each of these three HBM components that will hinder CXMT’s progress at different HBM generations. Each merits its own discussion.
DRAM
The memory industry uses a different terminology to mark node sizes compared to the logic industry. Instead of referring to a node by nanometer, the DRAM industry has begun to use letters for its advanced nodes. They started first with 1x, then 1y, and then 1z; afterward, they moved to the Greek alphabet, with 1α after 1z, and then 1β, and then 1γ. (Samsung and SK Hynix use the English 1a, 1b, and 1c instead, but this article uses Micron’s terminology.) Just to demonstrate the gap between each generation, between Micron’s 1β and 1γ nodes, the product speeds increased by 15% while reducing power usage by 20%.
As of 2025, CXMT is three generations behind the leading memory manufacturers, making the 1z node while the big three are shipping 1γ. With the 1z node, however, CXMT can produce DRAM for HBM up until HBM3.
But what must CXMT do to achieve beyond the 1z node? To get to 1α and beyond, CXMT must shrink DRAM cells even further, which requires advanced tools in lithography, etching, and deposition.
Lithography
Two of the most difficult steps in DRAM manufacturing are forming the bitline contact (BLC) and storage node contact (SNC). The BLC is the physical connection between periphery transistors that decide what memory needs to be fetched to amplify their signals and the capacitors that actually hold the memory.
As shown above, patterning and etching the BLC must thread the needle so as to contact the source/drain of the array transistors rather than the buried wordline (BWL) shown in teal.
The case is similar for the SNC, the physical connection between the bitline and capacitor. As shown below, the SNC must be etched through layers of different materials to again connect with the source/drain of the array transistors, instead of the BWL.
As DRAM nodes progress, the pattern density and critical dimensions of these processes get stricter, and greater precision is required. Eventually, EUV lithography is needed for these processes.
However, Micron has used techniques like self-aligned quadruple patterning (SAQP) to continue to use DUV up until its 1β node. Chinese manufacturer SMIC has used similar techniques to stretch DUV use for advanced nodes in the past, like its 7 nm Huawei chip. CXMT is likely even better at utilizing SAQP given the memory industry’s lengthier history with the process. Even for 1γ, Micron only uses EUV for one layer of the process, likely either the BLC or SNC step.
Thus, CXMT can likely also stretch its DUV use until 1β. After that, considering Micron has attempted to delay EUV use until the last possible moment, 1γ and beyond will become extremely difficult without access to the export-controlled EUV equipment. Without EUV, advanced nodes will either be impossible to make or of terrible yield; according to some estimates, using EUV, while more expensive, saves about 3-5% yield for advanced nodes while decreasing process steps by 20-30%. Without EUV, CXMT’s progress in DRAM will likely be stalled at the 1γ node, meaning HBM4E and beyond will be difficult for China to achieve from the DRAM standpoint alone.
Etching
For etching, the picture looks more favorable for CXMT. Advanced etching is required for the steps above, as well as for creating capacitor holes. These holes, which hold the memory charges, have small critical dimensions, high pattern density, and are very deep. Etching narrow yet deep holes like this can lead to a variety of defects, shown below, and thus require advanced tools with high aspect ratios (ratio of height to diameter). Aspect ratios reached 40:1 in the 1x era, with estimates for advanced nodes closer to 60:1.
The U.S. has imposed export controls on advanced etching equipment, including anisotropic etchers (the ones needed for capacitor etch), though China has been able to domestically produce equipment defying the controlled parameters.
For etching through silicon nitride for the capacitors, BLC, and SNC, Chinese products include Naura’s Accura NZ and Accura LX, as well as AMEC’s Primo nanova. Technical specifications about Chinese products are not widely available, though the Primo nanova is specifically advertised for the 1x node and beyond. Although this means the product probably cannot be stretched to cutting-edge nodes, Naura’s tools may work well enough.
Regardless, the existing Chinese offerings demonstrate that China is not too far behind on equipment for capacitor etch. These tools are susceptible to having exaggerated capabilities or scaling issues with manufacturing, but, especially compared to lithography, they’re not so far behind. China holds 10% of the global dry etch market and is self-reliant for about 15% of its advanced etching needs. The country’s rapid growth in the industry also demonstrates that etching obstacles may not be so solid. In short, China’s HBM progress will probably not be meaningfully hindered by DRAM etching bottlenecks.
Beyond etching, advanced deposition tools are required for DRAM manufacturing, but the story is very similar to etching: China can already produce the tools required, so it will likely not be a bottleneck. China is self-sufficient for 5-10% of its deposition needs and is also rapidly accelerating its indigenization efforts.
Through-Silicon Vias (TSVs)
Another step in DRAM manufacturing for HBM is the formation of through-silicon vias (TSVs), diagrammed below. This front-end-of-the-line process forms the vertical connections that allow stacked DRAM dies to communicate and function together. Without TSVs, the concept of HBM and of nearly all advanced packaging would be impossible.
For making TSVs, the most important process again is etching. TSVs require precise etching through DRAM dies to later deposit the material that serves as the vias connecting all the wafers together. The U.S. has imposed export controls on etching equipment specifically for TSV formation (EC 3B001.c.4), but again, China’s domestic manufacturers have been able to defy these parameters.
TSV critical dimensions currently range from 3-5 µm with depths of less than 100 µm. As nodes progress, DRAM dies are getting thinner, and both the depth and CD will decrease. Currently, China already offers equipment to satisfy these TSV requirements. AMEC’s TSV300E advertises a TSV CD of down to 1 µm and can achieve depths of several hundred microns. Naura’s PSE V300, though not publishing its specs, likely achieves a similar performance. Chinese product specs may be exaggerated or with lower throughput, but empirically, TSVs do not seem to pose an issue for CXMT given its capacity rivals other leading memory makers.
Having already achieved likely self-sufficient capabilities in TSV formation, CXMT will not be bottlenecked from this step in HBM manufacturing.
High-κ Metal Gate (HKMG)
Another process difficult in DRAM manufacturing is implementing the high-κ metal gate (HKMG). As shrinking DRAM cells for performance gains becomes increasingly difficult, HKMG has served as another means to increase device speeds.
As shown below, periphery transistors on a DRAM die are normally advanced by shrinking distances between the source and drain while also thinning the gate insulator. However, when insulator thinness reaches its limit, leakage issues emerge, and HKMG is used to solve them.
HKMG replaces traditional gate materials in periphery transistors to accelerate electron flow and prevent power leakage. Partially due to implementing HKMG, SK Hynix was able to achieve a 33% boost in speed with a 21% decrease in power usage.
The HKMG process has been adopted by memory makers since, and CXMT is now beginning its adoption process too; however, some reporting indicates that CXMT is struggling with its HKMG implementation, leading to reduced yield and slower manufacturing ramp-up. Other memory makers have adopted HKMG in their process flows around the 1z node, where CXMT is stuck now, so the company must hurdle the HKMG barrier to keep pace.
Incorporating HKMG in DRAM processes is difficult, partially because of the simultaneous processing of the periphery and array on a single wafer. The thermal budget of the array, or how much heat the structures are able to withstand, is relatively low; this means that the standard HKMG processes for logic nodes cannot be so replicable for DRAM. Although CXMT is currently struggling with HKMG, this doesn’t seem like an insurmountable issue. The bottleneck seems to be the more amorphous challenges of experimenting and perfecting process flows rather than a concrete wall of equipment inaccessibility. The equipment required for HKMG generally relates to the deposition tools in which China seems more or less self-sufficient.
Because of the lack of “hard” barriers like lack of access to tools, HKMG adoption will likely not be a serious hindrance to China’s HBM advances.
Base Die
The HBM DRAM dies sit on top of the base die. Among other functions, the base die routes signals coming in and out (I/O) of the memory stack. Ultimately, regardless of how strong the memory dies are, the power of the base die determines the upper limit of memory bandwidth for HBM.
As HBM nodes have progressed, the number of pins on the base die has increased, along with the data transfer speed of those pins. As a result, memory makers have used more advanced DRAM nodes to function for the base die to satisfy the requirement. Around the HBM4 generation, though, memory makers are compelled to use more expensive logic nodes to handle the workload. As such, memory makers are now partnering with TSMC to manufacture their base nodes for advanced generations.
The advanced logic nodes used for base dies will pose a problem for CXMT in its HBM advancement. Without EUV lithography, SMIC has been struggling to advance beyond 7 nm without abysmal yield.
For HBM4, CXMT can retrace Micron’s steps and continue to use a 1β DRAM die for base die functions. However, this decision would have significant drawbacks. Not all HBM4 are created equal, and by using a memory-process base die, Micron has emerged with HBM4 worse than SK Hynix and Samsung. While Micron’s product meets the JEDEC minimum of 8 Gbps per pin and goes to 9 Gbps, SK Hynix and Samsung have been able to reach 10 Gbps per pin and beyond via logic node base dies. Micron claims that they have begun sampling HBM4 with 11 Gbps, but Irrational Analysis explains why this is probably misleading.
Regardless, Micron has conceded that memory nodes are not best suited for the base die after HBM4 and has partnered with TSMC to produce the base die for HBM4E on an advanced logic node. For CXMT, this likely means that using 1β DRAM dies for HBM4 will result in a subpar product, and that HBM4E will be difficult to make without SMIC making breakthroughs in logic nodes.
However, lower cost HBM4 and 4E may be possible for CXMT. Although memory makers are producing their most advanced base dies for HBM4 at 5 nm and below, they are also offering alternatives with cheaper 12 nm base dies. 12 nm base dies can get the job done, but the products with more advanced logic offer smaller interconnect pitches for memory performance and lower power consumption. These make the 5 nm base dies attractive for AI workloads desired by customers like Nvidia.
Although CXMT could theoretically partner with TSMC for its base dies, as they would likely not fall under export control restrictions, my conversations with experts suggest that TSMC may not accept such orders given geopolitical tensions. Essentially, without access to advanced logic nodes for the base die, CXMT will likely struggle to make competitive HBM4 and HBM4E. They will likely be able to makeHBM4 with non-leading-edge 12 nm base dies. Perhaps they will even be able to secure orders from TSMC for advanced nodes, but the amount of question marks here makes CXMT’s success uncertain.
Packaging
Packaging is how the entire HBM stack comes together, and one element in particular is relevant. The “glue” that binds DRAM dies to each other, or bonding, is critically important. Stacking so many dies together creates thermal issues that bonding plays an important role in addressing; further, more efficient bonding with minimal gaps between dies is important to enable further stacking. As HBM has evolved from stacking only four dies to now up to sixteen, efficient bonding has been a key enabler.
Die Bonding
A possible struggle for CXMT will be succeeding in die bonding, but not because of export controls. Currently, export controls do not restrict the sale of bonding equipment used for HBM.
The two primary methods for die bonding in HBM are thermocompression bonding with non-conductive film (TC-NCF), used by Samsung and Micron, and mass reflow-molded underfill (MR-MUF), used by SK Hynix. SK Hynix adopted MR-MUF early on since HBM2E, and because of the decision, SK Hynix has been consistently lauded as creating superior HBM.
MR-MUF involves heating and connecting all the stacked dies at once, rather than one at a time like in TC-NCF. The real magic potion for MR-MUF, though, is the epoxy molding compound (EMC) used to fill the gap between dies.
MR-MUF has both better throughput and thermal dissipation than TCB. This is important both to scale production of HBM, but also to manage its heat requirements. By using MR-MUF, SK Hynix is able to stack more dies with fewer usage problems. HBM failures are the number one cause of AI chip failures, so MR-MUF to manage heat grants a real competitive edge.
Following SK Hynix’s footsteps, CXMT is reportedly adopting MR-MUF for its HBM3 and beyond; however, adoption is not like flicking a switch. To reap the benefits of MR-MUF, CXMT must solve several issues. First, MR-MUF is inferior to TC-NCF in managing die warpage. As DRAM dies become even thinner, CXMT will take time resolving this issue, just as SK Hynix has. SK Hynix solved this issue with a process it calls “advanced MR-MUF,” which adds a step of temporary bonding to the process — a step which CXMT may imitate.
Secondly, material acquisition may pose a problem. Competition, not export controls, may bar CXMT from acquiring the EMC for MR-MUF. SK Hynix has an exclusive deal with the Japanese materials company NAMICS for providing its EMC. SK Hynix’s material has been co-developed over years with NAMICS, and the material must be suited for each company’s process flow. Some Chinese sources suggest that CXMT’s EMC supplier is the domestic company Huahai Chengke (华海诚科), but this is still unconfirmed. Even if CXMT uses a domestic supplier, it will likely take years to work together to achieve a high yield.
Because of the extra steps from DRAM making to die bonding via MR-MUF, CXMT’s yield for its HBM3 in 2026 will likely take time to ramp up. Some experts claim that CXMT’s HBM3 yield likely won’t break 40% until the latter half of 2026, partially because of the MR-MUF adoption process.
In the end, though, CXMT’s early bet on MR-MUF will likely turn out to be a good idea in the long term, if not the short term. The advantages of the process are clear, and the bonding process only seems to be a short-term stumbling block. Though not a strict bottleneck, adopting MR-MUF will likely cause CXMT to slow production of HBM3 and beyond, but will not serve as a bottleneck for advanced generations.
Unanswered Questions
It is difficult to gauge CXMT’s capabilities or breakthroughs with 100% certainty. Unlike Chinese model developers, China’s chip manufacturers like to play their cards close to their chest. Because of the sensitive nature of their work, which is relevant for national security goals, or perhaps just because of the nature of the industry, CXMT rarely makes public statements. Perhaps this will change if CXMT undergoes its IPO as planned in 2026.
As such, certain details about China’s memory ecosystem are unanswerable without insider information. Some specific questions are listed below, and ChinaTalk invites anyone with color to reach out with answers or leads:
DRAM Node Sizes
What are the critical dimensions of the latest DRAM nodes and their aspect ratios?
What are the critical dimensions for TSVs in the latest HBM generations? How many TSVs are now included on a single DRAM die?
Chinese Equipment Ecosystem
How good are AMEC and Naura’s etching equipment for mass production? How good is China’s deposition equipment in practice? How true are the advertised specs?
CXMT Struggles
What part of HKMG adoption is CXMT struggling with?
Who is CXMT’s EMC provider for MR-MUF?
If anyone has answers to any of these questions, or has information related to prior analysis, please respond to this email or reach out to jordan@chinatalk.media!
Conclusion
Overall, CXMT is progressing at a steady pace for making HBM, but this trend is likely not to hold forever. For each step of the HBM process — DRAM, base die, and packaging — different bottlenecks will appear to stall CXMT’s progress or compel them to make sub-par HBM. First, the lack of advanced logic for base dies will likely lead CXMT to make lagging-edge HBM4. Even if CXMT utilized a memory node for its base die for HBM4, this would result in an estimated 10% decrease in memory bandwidth. After HBM4, both the base die constraint and the lack of EUV for DRAM manufacturing will cause trouble.
Summary of Conclusions:
But CXMT should not be written off. The industry chose HBM as the best option for memory in AI chips because it was the path of least resistance. With export controls, that may not be true for CXMT and China. Other alternatives for alleviating the memory bottleneck have been discussed, including using hybrid bonding, high-bandwidth flash (HBF), a unified cache manager (UCM), compute in memory (CIM), ferroelectric RAM (FeRAM), and magnetic RAM (MRAM). All of these options have their own problems and are nowhere near adoption, but they present opportunities for China to move off the beaten path and achieve memory self-sufficiency in its own way. If any U.S. administration reverses export controls, though, China will be able to more quickly follow the path for HBM development and catch up in the AI chip race.
For now, though, with HBM remaining the preeminent option, CXMT will have its work cut out for itself.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
, the new interim Executive Director of CSET who substacks at , makes her ChinaTalk debut. Present since the founding of CSET, Helen has had a front-row seat to the drama shaping today’s AI world — including a stint on OpenAI’s board.
Today our conversation covers…
What it means to run CSET in 2025, and how to keep think tank work rigorous and relevant in the age of AI,
The “good-faith” vs “dark arts” actors shaping Washington’s AI policy debate,
What her recent trip to China revealed about how Beijing is thinking (and not thinking) about AI,
Why AI progress might stay “jagged,” and what that means for AI policy,
The 2026 Tarbell Fellowship is now open. You could come work with us at ChinaTalk! Apply here.
Don’t just take it from me. Take it from our current Tarbell Fellow, , on his experience so far:
“Tarbell placed me at ChinaTalk for a year, fully funded! It’s been a dream setup to report seriously on China, tech, and AI. The fellowship’s training covers both journalism and the fundamentals of AI, which makes it one of the best on-ramps for people who didn’t come up through traditional reporting or AI pathways.
I always thought about tech journalism but assumed I missed my chance after college. Tarbell gave me another shot. ChinaTalk gives me the freedom to chase questions I’m genuinely curious about in the China–AI space, paired with a team that constantly reads each other’s work, shares articles, and brainstorms ideas. You’ll be producing impactful work for a large audience, but you’ll also be learning every day.
At ChinaTalk, I spend my time digging into the semiconductor supply chain, Chinese AI models, U.S.–China relations, and whatever else I get excited by. If that sounds like your idea of fun, apply!”
Think Tanks in the Age of AI
Jordan Schneider: As the new interim Executive Director of CSET, are you excited to rip up everything they’ve created and remake it in the image of Helen Toner? What is your vision for the future of CSET?
Helen Toner: If there’s one thing that I have learned from the many friends and colleagues who’ve rotated in and out of government, it’s that your day-one mission should be reorganization. Step in, tear everything up, and change the org structure.
No, I’m kidding. It’s exciting and an honor to be in this position. After Jason and Dewey, I’m stepping into big shoes. I’ve been at CSET since its founding in 2019, so it’s exciting to shepherd the organization into a new phase.
CSET’s success is built on a foundation of excellent work, and I want to continue that. The core of our mission is to produce intellectually independent research that is driven by evidence and data.” Our data science team is unique in the think tank world — their data powers our analysis. On every project, we make sure our analysis is rigorous and driven by the best evidence we can find. We care that our work is technically informed.
One of our founding goals at CSET was to show a different way for think tanks to operate, and ideally, inspire others to follow us. I think we’ve been really successful there. You now see RAND with a huge emerging tech and national security effort, and CSIS doing more translations and data visualizations — things that were core to the CSET model and are now much more common in Washington.
That’s great, because it proves our model works. Of course, it also means we have competition, so we have to show what makes CSET unique and where we provide particular value. Our deep expertise on China is a perfect example. We have a whole range of China specialists woven throughout our team, covering everything from language to specific subject matter. I’m excited to lean into that and to keep evolving. Emerging tech never stands still, so we have to keep figuring out where we can add the most value.
Jordan Schneider: I agree that CSET has raised the bar for discourse in Washington — it’s why I gave CSET ChinaTalk’s only “Think Tank of the Year” award back in 2022. It’s been heartwarming to see your standard of using real evidence on thorny topics like chip controls, immigration policy, or the PLA’s use of AI resonate so strongly in the broader debate in Washington.
But at the same time, we’re seeing a paradox. Since 2019, it feels like facts matter less than ever. Arguments get reduced to tweet-shouting matches, and remarkably, those shouting matches are now becoming central to the actual policy debate on AI. What’s your take on these two trends happening in parallel? What’s the synthesis?
Helen Toner: I think there are multiple layers here. You have the headlines in the New York Times or the Wall Street Journal, but there is also work happening beneath the surface. The U.S. government has millions of employees, and the subject matter experts doing the work are interested in details and evidence. There’s a steady demand from them for the kind of support we provide, and they are very responsive to facts.
Another example is the discourse around recent AI legislation. Take California, for example. Last year, the discourse around the SB 1047 bill was awful. Then this year, they convened a governor’s panel, published a report, adopted its recommendations, and passed a less controversial bill. It’s a crazy turnaround. We saw something similar with the EU Code of Practice — it looked like it was going to fall apart, but then it came together. I don’t want to sound too pollyannaish, there’s a lot to be concerned about. But it’s important to remember that sensible work is still getting done.
Jordan Schneider: I started ChinaTalk in 2017, and CSET started in 2019. Back then, the intersection of U.S.-China relations, emerging technology, and national security was not a front-page topic.
Helen Toner: When we said we wanted to have a whole organization focused on emerging tech and national security, and people were like, “A whole organization? Like, four people?”
Jordan Schneider: It’s been a wild adjustment for this space to go from an idea funders would laugh at to something presidents tweet about all the time. But that shift has also brought in layers of bad faith. Back when this community was smaller, there weren’t many people playing dirty.
I think CSET has its heart in the right place and is doing earnest, yeoman’s work. But there are snakes in the grass everywhere now. There’s so much money riding on this research, and that wasn’t true a few years ago. I admire your pollyannaishness — I think it’s good for your mental health. But is the most effective option to put out good research and facts? Or are “dark arts” needed to have that research shape policy?
Helen Toner: I don’t think the only options are “put a white paper on your website” or “go full political dark arts.” There’s a lot of space in between. From the beginning, we’ve done more than publish research — we actively seek out the relevant policymakers, brief them, and work with their teams on legislation. Now, we’re also thinking about how the internet has changed what that means for us. Should we be doing videos? I’m not sure, but we should at least consider it.
Another big shift, which I know you follow, is the trend toward individual brands over institutional ones. Some of our people are eager to give that a go, while others — especially those from the intelligence community — are like, “Oh God, shoot me before you make me go on Twitter.” We’re exploring that space — finding ways to keep doing good-faith, fact-based work while operating effectively in today’s ecosystem.
Jordan Schneider: I worry that an organization where good-faith, facts-focused people are comfortable is fundamentally different from one with “dark arts” specialists. The cultures and incentives don’t mix.
Helen Toner: Will there be a ChinaTalk “Dark Arts Think Tank Award”? Who would that go to?
Jordan Schneider: Wow, I don’t know. I can’t give any names here — this is for public consumption. But I agree that there will always be an audience for grounded data, and someone needs to provide the facts.
Helen Toner: When we talk about “the Facts,” it’s not about some ideas being more virtuous than others. But if you want to accomplish something and care about results, then you need to know what the world looks like.
We worked closely with the Biden administration when they were considering outbound investment controls — asking them, “How will you implement these controls? Do you have the necessary information to do it effectively?” This isn’t about taking a holier-than-thou position. It’s about the reality that if you don’t know what’s going on, you’re going to try things that backfire — and most people want to avoid that.
Jordan Schneider: A better framing might be that it’s better to have data in the discussion.
It’s remarkable what a single researcher can do in recent years with an “individual brand”. It’s wild to think that CSET was around before we could ask ChatGPT what The PLA Daily 解放军报 says. I can now code data visualizations in two hours, which I used to assume would require a CSET-level team and budget. How do you think these new tools change what a solo researcher or a small team can accomplish? Does this change how CSET operates?
Helen Toner: We’re looking at it from the opposite side — what unique things can our larger team do that an individual still can’t? Our data team is tackling a huge data science problem called “entity resolution”. That’s figuring out that “Google London” and “DeepMind” are both “Google” in a massive dataset of text. It’s a huge, messy problem, and using language models in a carefully designed and validated pipeline, we blew past previous results.
We also analyzed ~3,000 AI contracts the PLA is buying and used language models to parse that data. As a larger team, we can do things an individual researcher can’t. We can validate our results and test different models — like when to use an expensive, frontier model versus a lighter one that’s faster and can handle high volumes. We’re doing tons of experimentation there, and the team is coming up with some really cool stuff.
Jordan Schneider: CSET was early on important AI topics, but has remained ideologically neutral — you were writing about semiconductor export controls in 2020, but have not published an “AI will arrive in 2027” style analysis. Is now the time? Are the odds of those radical changes high enough that you need to start spending your team’s time and budget exploring them?
Helen Toner: I have a unique perspective because I have my feet in two worlds. I’m from the AI safety community, which is in that mindset, but most of my team at CSET is not “AGI-pilled.” We’ve done a lot of work on scaling and red-teaming, but not the “OMG AGI” work. We are currently hiring someone to work on frontier AI issues, and I’m hoping to increase our work in that space.
Jordan Schneider: What are you excited for this person to do?
Helen Toner: I’m excited about the “Frontier AI” framing. I’m glad RAND is now researching AGI, but the concept of AGI is messy and contested — it’s not clear what, if anything, is there. There has been a giant gulf between the AI systems we have — those we can touch and test — and hypothetical concerns about future AGI. But in two years, that gulf has gotten smaller. Now we can look at current systems and extrapolate future ones — which makes this topic amenable to CSET’s evidence-based methods.
I’m psyched for this research. It won’t be “CSET predicts AGI in 2027,” but it’s important to consider the possibility of AGI or superintelligence on timescales soon enough to matter for policy. Watch this space.
The Jagged Frontier
Jordan Schneider: You’ve expressed the view that AI progress could stay jagged. Can you elaborate on that?
Helen Toner: The idea comes from Prof. Ethan Mollick, and possibly also Andrej Karpathy. I highly recommend following Mollick’s AI work on Substack, LinkedIn, or Twitter. His idea was a “jagged frontier” — that AI is good at some tasks and surprisingly bad at others.
I recently gave a talk on this, arguing we should take seriously the idea that AI’s progress might remain jagged. Right now, most people fall into one of two camps — either they think AI is all hype and a “nothing burger” — or they’re in the “AGI by 2027” camp. That group believes powerful AI will become a drop-in remote worker or an automated AI engineer. Both are non-jagged views of the future. The question of what persistent jaggedness would look like is underexplored.
Jordan Schneider: The “jagged frontier” idea is more nuanced than mainstream discourse on AI — the Twitter brain, swinging wildly between “it’s over” and “we’re so back.” Why do you think people resist the possibility of uneven AI development — that the next model won’t solve everything? Why does the jagged idea struggle to gain traction, even though it is our current reality?
Helen Toner: Most people agree today’s AI is jagged, but they believe the future will be different. I think that’s because we use humans as a reference point — we believe that what’s difficult for us must be universally difficult, instead of seeing it as a product of our own evolution. Since the 1950s, we’ve debated — are we recreating the human mind, or building useful machines?
We’re currently far down the “build useful machines” path, but the idea of recreating the human mind is built into the field. I think this is why people expect AI to be more human-like than it is.
Jordan Schneider: There’s money involved now — the AI hype is backed by enormous financial incentives.
Helen Toner: Jaggedness doesn’t only refer to the troughs where AI struggles — there are high peaks as well. In the next 5-10 years, I expect us to exploit the heck out of those peaks. But the way we do so must account for the troughs.
Jordan Schneider: Is jagged AI more tractable for policy research? Is CSET’s work more relevant in that scenario?
Helen Toner: If jaggedness persists, fast takeoff scenarios are less likely — scenarios like an automated AI researcher that makes ten years of progress in six months. That would be a hard world for policy to operate in — there isn’t time for the government to form a commission, write a nice report, and debate it in the next legislative session. Jaggedness leads to slower AI progression, which gives us time to reflect, experiment, and adapt.
I’m not certain jaggedness will persist, but the idea is underrated in the AI community. At the same time, we should consider the possibility of non-jagged, rapid AI progress. It could still happen, although it’s not my best guess.
Jordan Schneider: There’s a resource allocation problem in AI policy research. Should we focus on a tangible, near-term jagged frontier — like AI’s impact on cybersecurity — or on the sci-fi futures of self-improving AI? People are drawn to speculative, sci-fi scenarios — a cybersecurity paper won’t go viral like “AGI by 2027” did. But there is value in working on a more probable future.
Helen Toner: There is a lot of low-hanging fruit in research on jagged development, and a lot of possible futures. What will AI be good at? What tasks will it struggle with? What does that mean for adoption and integration?
A jagged frontier means we are unlikely to fully automate complex jobs or goals. Instead, we will get powerful AI advisors and a “centaur” model of human-AI teaming, which you mentioned in the AI girlfriends podcast. Future human-AI collaboration scenarios are underexplored because predictions of super-powerful AI assume everything will be automated. They focus on abstract problems like alignment, not the messy, practical details of human-machine teaming that a jagged world would demand.
Jordan Schneider: After writing a paper on AI honeypot espionage, I decided to do some experimenting. Over the past few days, I’ve tried to fall in love with an AI, and it’s not lovable in the slightest.
When it comes to personal comfort and consolation, AI jaggedness is very apparent. There has been a lot of recent reporting about people who’ve developed close, intimate relationships with AI, but it’s not doing it for me. What should I make of that, Helen?
Helen Toner: Have you tried the Grok anime goth girl? You need to find the right one for you.
Jordan Schneider: It was bad — really repulsive. Even if I’m not the target audience for these AIs, if they were smart, they would have figured me out after 10 minutes of conversation — the way TikTok figured me out after 45 seconds of swiping. These models cannot do that — that’s an important detail.
The lack of personalized learning is a huge hurdle for AI in the workplace. Instead of learning from user input, models are trained and dropped into organizations, leaving people to figure them out. If the future of this technology depends on personalization that fits like a glove — professionally and personally — then we need to solve this.
Helen Toner: There’s a long way to go. We held a workshop in July about automating AI R&D and the potential for an “intelligence explosion” takeoff. We need to question underlying assumptions — what does progress look like? What are the gaps? How soon can we fill them? We’ll examine this in an upcoming CSET paper.
Jordan Schneider: There’s tension in our view of AI’s capabilities. It’s easy to overlook its limitations in work you do not do yourself, but in your own work, you can feel the jaggedness firsthand. You have an intuitive sense of where AI is exceptional and where it’s uneven.
Ironically, AI engineers are the most optimistic about AI’s capabilities — maybe a little high on their own supply. But the proof is in their paychecks — companies are hiring them in droves because AI cannot do their jobs.
Helen Toner: There are many sources of jaggedness.
A key source of AI’s jaggedness is the context window — how easy is it to input the organizational or practical context of a task? Some professions, like software engineering or marketing, are easily digestible for an AI because you can copy-paste the relevant code or creative brief. But most jobs can’t be reduced to a text file — their context is messy and organizational. We haven’t fully grasped how this single limitation will shape what AI can do and how quickly it can do it.
AI Debates in China
Jordan Schneider: Helen, you were in China recently. How was that trip?
Helen Toner: It was great to be back in China. In 2018, I was in Beijing for 9 months, studying Mandarin and learning about China’s AI ecosystem. But between my green card, the pandemic, and having kids, it had been ages since I was there. I went for a quick five-day trip to Shanghai for the World AI Conference, which was gigantic. You know what Chinese conferences are like — the huge stage, the flashing lights. Robots were walking around everywhere, something you couldn’t get away with in the U.S. Kids were petting little quadruped robots that were roaming the floor. It was a good time.
A robot dog display at the 2025 Shanghai World AI Conference. Source.
Jordan Schneider: Were you recognized?
Helen Toner: No, definitely not. Not that anyone told me.
Jordan Schneider: What’s your sense of the U.S.-China AI dialogue and opportunities for discourse or cooperation?
Helen Toner: People in the AI safety community often ask why there isn’t a U.S.-China dialogue on avoiding a race to superintelligence. The answer is that there is no agreement on what the problem is, or what the U.S. and China’s interests are. At a Chatham House discussion I recently attended, the Chinese organizers were divided on whether to focus only on superintelligence or broader development questions as well. Within their team, there was no consensus on the core issues. These conversations are a good start, but we still have a long way to go.
Jordan Schneider: A core AI policy question is how the U.S. and Chinese ecosystems will relate to each other. What are the other key questions that will define the field for years to come?
Helen Toner: On the national security side, the U.S.-China dynamic is a big one, covering both competition and the potential cooperation on AI. Military integration is another huge question. The focus is shifting from developing advanced AI to how it changes a military’s operational concepts and the way it fights. This is an adoption challenge.
There are also serious risks around cyber and biosecurity, but we might get lucky, and the threats are manageable. I’m personally more concerned about cyber, but I know well-informed people with access to classified information who are deeply worried about the bio risks.
Outside of national security, we’ll see more community-level issues, particularly around data centers. A narrative about their water use is gaining traction, and while the data may not show an unusual amount of consumption, the community perception is strong enough to create backlash. There are also social questions. We do not have a framework for dealing with AI companions, especially for children, and the impact of AI on labor and jobs is not going away.
AI Parenting Advice
Jordan Schneider: Do you have any AI parenting takes, Helen?
Helen Toner: I have a three-year-old and a one-year-old, so thankfully, we’re not there yet. But I worry the “engineer-brained” approach to parenting reduces child-rearing to a set of tasks. The idea that if an AI can entertain or teach a child “better” than a human, then it’s a net win, misses the point. The relationship between a parent or teacher and a child is a huge part of what it means to grow up and learn. AI should be a tool to enhance connection, not replace it. If an AI generates a story, read it to your child, but do not be too utilitarian. What are your thoughts?
Jordan Schneider: Abstracting love is a high bar for AI.
Kids are wired by billions of years of evolution to trust a warm, sweaty mammal. An AI can certainly teach them physics or math better than I can, and outsourcing that is one thing. But the biological need for connection is another. Primate studies show the same thing — the monkeys want to be held. Trying to engineer that need away is playing with fire. Maybe a robot will get there in 20 years, but you’re running hard against evolution. No offense to anyone using Midjourney for children’s books — I have that tab open right now.
Japanese snow monkeys embrace in the cold. Source.
Helen Toner: I think there are good ways to do it.
Jordan Schneider: Absolutely. But the sci-fi future where kids don’t need loving parents for connection or as models of how to relate to other humans seems a long way off.
Helen Toner: There is a This American Life story that sticks with me, about a single dad and his daughter. He was a physicist, and she would ask him astronomy questions like, “Why do stars...?” or “Where did the Earth come from?”. Kids love to ask “why” questions. He found answering them stressful, so one day, he asked her to write down all of her questions. He locked himself in his office and wrote up a gigantic set of answers for her. The interviewer on the show asked the girl what she thought, and she said, “I wanted to hang out with my dad.” It’s so tragic. Don’t do that with AI.
On Calling Timeout
Jordan Schneider: My theory is that CSET only exists because of Jason Matheny. The national security risks of China’s rapid AI growth were completely off the radar for these funders. It took an exceptional person they trusted, like Jason, to convince them to build a community around this idea.
Before CSET, there was no tech team with deep China expertise. I spent years trying to make the case that competition with China mattered, that AI was more than one small piece of a larger puzzle, but people were unconvinced — that idea was ’too spicy’ or too far out.
There was a brief moment during the 1st Trump administration when it became a mainstream concern. Many corporate blogs, including that famous OpenAI document, were suddenly about beating China. But that moment has passed, and it feels like the issue is becoming less relevant again.
Helen Toner: It’s an interesting time for China+AI policy. When I started in this space around 2017, people in AI would ask, “Why talk about an AI race with China?” and then give AI-specific reasons why it wasn’t a race. I had to explain that they were missing the bigger picture. The U.S. national security apparatus was orienting towards strategic competition with China. For them, AI was only one small manifestation of that competition, and the AI community’s arguments were seen as irrelevant noise.
Jordan Schneider: I remember people telling me, “Oh, but if we say this, will it accelerate the race?” Bro, come on.
Helen Toner: Is strategic competition with China still the main goal of the U.S. national security apparatus? People outside the tech world are not sure — the current U.S. policy toward China is unclear. That’s disconcerting in some ways, but it also creates potentially productive space.
Jordan Schneider: In the 1st Trump administration, U.S.-China competition was a central pillar — Jake Sullivan wanted “as large a lead as possible.” But in Trump 2.0, focus on China has waned, and now we’re mobilizing against Venezuela while ignoring Chinese boats in the Philippines.
For years, I’ve thought that the U.S.-China AI race was inevitable. In many ways, it has already happened — we now have bifurcated ecosystems for AI chips, models, and hyperscalers. These dynamics seem resilient to the day-to-day whims of policymakers. How durable is this rivalry? If the American president is not focused on this issue, do the competitive dynamics of the last decade have enough momentum to continue on their own?
Helen Toner: I don’t know how resilient the rivalry will be. The competition was never about AI — competition with China was the organizing principle of the U.S. national security apparatus, and AI was one part of that. The U.S. AI sector now uses that narrative as justification for everything from faster data center permits to avoiding AI regulation. If those arguments lose force, I’m not sure what will happen.
My own prediction has always been that China’s internal demographic and economic challenges would eventually cool the rivalry, though I thought it would take longer, maybe till the 2030s. With a president who is hard to predict and an increasingly isolationist MAGA base, and a new focus on the Western hemisphere, disengagement with China could be stickier.
Helen Toner: Hilarious episode in the sitcom that we live in.
Jordan Schneider: Exactly! This was a bill Congress passed almost unanimously, and then the president decided he was not concerned with an issue Congress had a bipartisan consensus on — that is an interesting detail. I’m not sure how illustrative that is.
Helen Toner: Another source of tension in the Trump coalition right now is between the “tech right” and MAGA. They have disagreements about whether to charge ahead with AI — whether AI is the best thing since sliced bread, or the devil, or the Antichrist. There is a lot of division, but both sides are less concerned about competition with China. The tech right wants to sell to China, and the MAGA world would prefer to slow the rate of development.
Jordan Schneider: Corporate self-interest could be a reinforcing driver. U.S. firms do not want to compete with Chinese companies on their home turf. Once Chinese EVs start taking market share, or Huawei chips threaten Nvidia, the game changes. The question will become, is access to China’s market worth giving up our own? The likely answer is no. U.S. companies will demand the same protected home base that Chinese firms have used to their advantage, which only accelerates the competition.
Helen Toner: Wait, they can operate here? I thought it was a one-way street.
Jordan Schneider: Bill Bishop often invokes a Xi Jinping quote that essentially says, “Our goal is to become more self-reliant at home and make the world more dependent on us.” This mindset was in the rare earths saga, where China’s escalation was a self-inflicted wound. It showed a willingness to compete in a way that alienates American elites. You can admire their ambition, but the U.S. will not accept Chinese competitors dominating key verticals — especially in the tech sector that underpins the U.S. stock market.
Helen Toner: We’ll see. I do not know how we can ban Chinese open-source models, which I think is one of the biggest threats to U.S. market share. Using open-source Chinese models presumably displaces API market share for OpenAI, Anthropic, or Google.
Jordan Schneider: They are not un-bannable — stopping individual downloads of Chinese software is a fool’s errand, but that’s not the real game. The real game is preventing billion-dollar companies from being built on Chinese open-source models, and the government has plenty of ways to do that. They can block Chinese models from government contracts, tie it up in FCC compliance issues, or make it a mandatory risk disclosure. If the U.S. government really puts its back into it, it can find a way.
Helen Toner: The government procurement restriction is a good point. Public company disclosures — that’s interesting. I agree, these policies can make it harder.
Jordan Schneider: Or change the incentives. Switching tangents, can you pitch some of the best recent work from CSET — what do you admire and plan to build on?
Helen Toner: One of our most exciting new papers analyzes 2,800 PLA AI contracts. The initial piece focuses on who is buying, and the key finding is that while the largest contracts go to state-owned enterprises, the bulk are awarded to “non-traditional” private companies and universities. More research is coming on what they’re buying.
Our work on DoD AI integration has also been impactful. Interestingly, our research has been valuable to government officials because it is public. Internal reports are often classified and hard to share, so a URL they can circulate is a game-changer. Our paper “Building the Tech Coalition,” which analyzes their use of Project Maven and the internal talent required, is a great example of this.
Number three is our work on AI and biorisks. The debate has been narrowly focused on controlling AI models, so our “Toolkit for Managing Biorisks from AI” broadens the conversation by outlining a full range of policy options, which has been helpful for policymakers.
Jordan Schneider: Let’s do two more, oldies but goodies.
Helen Toner: For oldies but goodies, I’d point to our outbound investment work, where we asked the Biden administration, “If you want to control outgoing investment, do you know how to do that? What data do you have, and what data would you need?” That implementation was a classic example of our work.
Our explainers have been surprisingly impactful. We published one about the differences between generative AI, large language models, and foundational models. A government agency was trying to decide which terminology to use in an influential policy document, and told us the explainer directly influenced their policy. Straightforward research like that has a good track record.
Jordan Schneider: Would you like to recommend some mood music to end the episode?
Helen Toner: The great China and AI scholar, Matt Sheehan, told me instrumental playlists are the best way to focus, so I’ve been listening to a lot of instrumental music. There’s a great James Brown instrumental album. Why not some instrumental James Brown?
Login
