After the Nvidia scandal, it seems clear the United States will act to constrain China’s access to AI hardware. The prospect of tighter export controls naturally raises two questions about what comes next: How long until China catches up in its own chipmaking? And—have we bought enough time to voluntarily slow AI capabilities research?
Since 2021, I have written on various dimensions of America’s national debate over semiconductor export controls. Early CSET research established that the Chinese military is wholly reliant on U.S.-designed computer chips; and weighed the financial impact to the U.S. semiconductor industry if we were to curtail that access. My more recent work has attempted to balance the wisdom of a global AI exports program with the need for China-specific export controls. Colleagues and I have closely tracked the pace of China’s chip industry development, its dependence on imported lithography equipment, and measures the United States might take to blunt Chinese chip production.
I think stronger chip export controls are the right policy to slow China’s capacity to build and deploy frontier AI. But policymakers should also understand that China could catch up faster than many of us would like.
How Long Until China Becomes Self-Sufficient in AI Hardware?
Suppose the United States succeeds in restricting China’s access to computing hardware. How long will it take Huawei and the rest of China’s semiconductor industry to manufacture high-quality AI accelerators at a scale sufficient to meet domestic demand—and begin competing in global markets?
As I wrote a few months ago with Satvik Pendyala, there are four variables worth tracking closely:
The first is advanced logic fabrication yields. China’s ability to manufacture sophisticated processors depends not just on having the necessary fabrication equipment, but on the portion of usable chips produced from each wafer.
Even minor improvements in manufacturing efficiency could significantly increase the output of logic die needed to manufacture AI accelerators. If SMIC succeeds in doubling the share of usable dies over the next few years (through better process control and fewer defects), it will be able to produce many more AI accelerators, without necessarily expanding chip fabrication capacity.
The second variable is how China allocates advanced logic chip production. Huawei and its manufacturing partners face competing demands to make smartphones, consumer electronics, and AI accelerators used in large data centers. So far, the company has been devoting most of its scarce logic chip fabrication capacity to Kirin smartphone processors rather than Ascend accelerators, because they are harder to manufacture profitably.
In the wake of American export restrictions, this could easily change. A decision to prioritize AI compute over consumer electronics could drastically increase the usable supply of accelerator dies—again, without necessarily expanding fabrication capacity.
The third variable is high-bandwidth memory. Modern AI accelerators require enormous quantities of memory capable of storing and moving data—all those trillions of parameters of model weights. China’s domestic memory industry, particularly CXMT, is rapidly seeking to scale up production.
This is partly a manufacturing challenge, but also a question of equipment and financing. One issue worth watching is whether Chinese memory manufacturers can secure enough foreign customers to finance a rapid expansion of China’s domestic memory production. This would inject massive liquidity into China’s memory chipmaking industry, and scale China’s ability to make high-end AI chips in the years ahead.
The fourth variable, perhaps least understood, is advanced packaging. Manufacturing logic dies and memory stacks is not enough—they must be integrated into functioning accelerators through sophisticated packaging processes.
Chinese firms including Tongfu Microelectronics are attempting to develop advanced packaging capacity at scale, to compete with TSMC’s Chip-on-Wafer-on-Substrate process. I have not seen a comprehensive public assessment of how many high-end accelerators China’s packaging industry can currently process. This may prove to be an important bottleneck on China’s domestic accelerator output for the foreseeable future.
Our provisional estimate is that China will begin manufacturing several million Ascend-class accelerators annually within the next three years, up from 1.3 million Ascends in 2026. Manufacturing at this scale represents only a fraction—between 1/8th to 1/4th—of Nvidia’s annual Blackwell output fabbed in Taiwan.
China’s indigenous chip output will still lag in compute performance. China’s ~1.3 million annually-produced Huawei Ascend processors will not be capable of delivering the same useful AI training or inference performance as 1.3 million of Nvidia’s comparable products. Nvidia-designed chips are fabbed using TSMC’s advanced manufacturing process, which achieves truly tightly-packed transistor density. Cut off from advanced lithography equipment, China must rely on new, unproven architectures to simulate that density—architectures that will struggle with heat and yield. For this reason, Nvidia hardware still outperforms Huawei’s in terms of FLOPS-per-chip, at a ratio of 4.5-to-1. Huawei’s chips need to draw more power to achieve the same performance. This is particularly important as power, not silicon, starts to become the binding constraint for nations wanting to reap AI’s strategic rewards.
Nevertheless, I think China has plausible pathways to producing large quantities of pretty-capable AI hardware within the next several years. What will matter most is that the United States can manufacture even more of it. The contest to deploy AI services increasingly resembles an industrial systems competition. Policymakers need to think and speak in terms of the national-scale production of devices capable of converting electricity into tokens.
What does all this mean for China’s actual AI capabilities? If the United States can succeed in blunting China’s compute access for 3-5 years, or even longer, this will have a material effect on the shape of AI competition between the two countries. To understand why, you should read Epoch AI’s October report, “How many AI agents could run on the AI chips shipped through 2027?”—to get a sense of how much agentic concurrency all this compute is buying us.
Epoch finds that AI chips shipped through 2027 could run tens to hundreds of millions of concurrent frontier-model agents. Running nonstop, they estimate these agents could supply as many weekly working hours as about 140–720 million full-time employees.
Blunting China’s access to compute could significantly reduce the number of concurrent frontier-model agents its enterprises are capable of running in parallel—significantly slowing its AI capabilities research, along with progress toward other strategic tasks.
If We Restrict China’s Access to AI Hardware, Can We Afford to Pace the Frontier in Software?
This question is complex because AI progress is not a single dial. I think it is useful to distinguish between two axes of progress.
First is the race to build more impressive frontier reasoning capability.
This is where American AI labs have a semi-durable advantage—if they choose to press it. The frontier increasingly involves large-scale training, reinforcement learning, and available test-time compute for extended and multi-agent reasoning.
America’s compute advantage allows models to spend more computational resources working through difficult problems before arriving at an answer. Scaling up reasoning capability is especially important for scientific research, cybersecurity, and building AI systems that can recursively self-improve.
A good share of Chinese progress in frontier reasoning appears to derive from distilling the reasoning outputs of U.S. models: taking Claude or ChatGPT’s chain-of-thought, and using it to train Kimi or Qwen.
There is good reason to believe America’s advantage in frontier reasoning will persist and even widen with the advent of RSI—especially if U.S. labs slow their public releases, or enact better technical guardrails against distillation. Many Chinese commentators and industry experts also believe this to be the case.
However, a slew of public safety events has prompted U.S. AI researchers to consider slowing the pace of their own capability development. Because compute is so central in unlocking frontier reasoning capability, one in-vogue proposal is to statutorily limit the amount of compute labs can dedicate to AI research activities (as opposed to running existing models—providing inference services).
Is this a good idea? On the one hand, America’s commanding lead in frontier reasoning would seem to imply that U.S. labs can get away with unilateral “pacing”—and still keep Chinese labs water-skiing behind the frontier. On the other hand, there is some logical inconsistency in simultaneously advocating that the United States 1) constrain the amount of compute available to Chinese labs, while 2) also constraining the amount of compute available for U.S. labs to conduct AI research. Some equilibrium between these positions is surely possible, but getting the balance right will be difficult—and getting it wrong could be costly.
A sufficiently large American lead in frontier reasoning might permit some deliberate slowing. But if efficiency gains, architectural breakthroughs, or new paradigms move the frontier without comparable increases in compute, American policymakers could badly misjudge how much strategic breathing room the United States actually has available to pace.
Second is the race to unlock architectural innovation and algorithmic efficiency.
Here I am much less confident that the United States will maintain a commanding lead, and I don’t believe architectural research should be paused outright.
China’s labs continue to find ways to do more with less. Releases from DeepSeek, Moonshot, Qwen, and others have experimented with sparse attention, memory management, and agent swarms. For example, DeepSeek’s September release, V4.1-Flash, reportedly reduced high bandwidth memory requirements for its key-value cache by 75 percent relative to the preceding model generation—a truly massive efficiency saving.
Architectural improvements that allow models to use less memory or process longer contexts more efficiently will increase the amount of useful AI work that China, as a nation, is capable of performing on a given stock of computing hardware. And, such advances can compound: Cheaper inference makes it possible to run more agents, conduct more experiments, and spend more computation on especially difficult problems—such as coding and AI research.
Some architectural decisions can improve performance at the expense of monitorability. Expanding a model’s “recurrent depth”—shifting more reasoning away from an explicit chain-of-thought, and into latent space—is one example of such a tradeoff. OpenAI caught some heat for experimenting with this technique in Astra, but the wider American AI industry norm against looped transformers has generally held. Unfortunately, I could easily see a future where Chinese labs come under pressure to cut corners on monitorability in exchange for improving model performance. As things stand, I don’t see an easy way to solve this coordination problem.
With low confidence, my mental model for where the future is headed looks something like this:
Governments fail to completely contain technological progress toward RSI—but they at least succeed in slowing both U.S. and Chinese AI capability development during a critical few-year window. The United States (mostly, temporarily) succeeds in cutting off Chinese compute, hamstringing both China’s research progress and inference capacity—while also voluntarily limiting the compute available for U.S. labs to conduct their own automated R&D.
Both the United States and China remain haunted by mutual dependence. For a few years, Washington and Beijing intermittently provide one another at least some advanced compute and critical minerals in fits and starts. Chinese labs face pressure to eschew monitorability in favor of capability development. China’s chipmaking kicks into high gear, with SMIC, CXMT, and Tongfu Microelectronics sprinting to expand logic, memory, and packaging capacity. Meanwhile, other sovereigns begin building multi-trillion-parameter language models of their own.
During this time, monitorability and alignment research improve; society hardens to withstand AI’s most disruptive effects; and the American technology stack becomes the global default for providing cognition-as-a-service.
The most important implication in this scenario is that U.S. hardware restrictions work, and American AI labs also deliberately slow some forms of capability development—but Chinese labs continue making progress through efficiency improvements that are harder to monitor or restrict.
Buying Time is Not the Same as Winning
I come away from all of this with mixed feelings.
On the one hand, I am encouraged that Washington appears willing, at long last, to restrict China’s access to advanced compute. It seems clear that doing so could materially slow China’s AI progress.
But I am not as confident that compute controls alone will be enough to cement a durable American lead across the full range of AI capabilities—especially if the United States does pursue a compute-based “pacing” regime of its own.
China has a sophisticated technology industry with capable researchers. Its semiconductor industry faces serious but ultimately surmountable obstacles. If compute capacity and agentic concurrency prove as valuable as many AI policy researchers expect, China will find a way to obtain them.
My experience in government has also taught me that American policy tends to move in response to incidents. We rarely anticipate the full consequences of a new technology before something goes wrong. Policymakers instead confront a handful of imperfect options, each with serious costs, and select the one they believe will best protect the American people against a narrowly scoped threat vector.
I expect AI policy to evolve in much the same way, particularly as governments begin to appreciate the national security implications of increasingly autonomous systems.
For now, there is an enormous amount of research to be done. We need better estimates of Chinese fabrication yields, memory production, and advanced packaging capacity. We need to understand how much computing power Chinese AI laboratories can actually access, even after the tightening of American export controls.
I think stronger semiconductor controls are the right policy—and I also think China could catch up faster than many of us would like.


