US export controls on AI chips may buy time against China, but a former policy analyst warns the gap could close sooner than Washington hopes. Tighter limits on China’s access to AI hardware look likely after a recent Nvidia scandal, and the open question is how long China needs to build its own supply.
The same analyst asks a second question that matters just as much: if the US cuts off Chinese compute, can American labs afford to slow their own research?
China’s four chip bottlenecks
Four variables decide how quickly China can build accelerators at scale, according to the analysis.
- Logic yields. If SMIC doubles the share of usable dies per wafer, it can make far more accelerators without adding fab capacity.
- Allocation. Huawei has put most scarce logic capacity into Kirin phone processors, which are easier to make profitably, rather than Ascend accelerators. US restrictions could flip that choice.
- High-bandwidth memory. CXMT is racing to scale up. Whether it can win enough foreign customers to finance expansion is worth watching.
- Advanced packaging. Firms such as Tongfu Microelectronics are trying to match TSMC’s Chip-on-Wafer-on-Substrate process, and there is no comprehensive public estimate of how many accelerators China can package today.

How many chips, and how good
The provisional estimate: China will produce several million Ascend-class accelerators a year within three years, up from 1.3 million in 2026. That is only one-eighth to one-quarter of Nvidia’s annual Blackwell output made in Taiwan.
Those chips will also be weaker. Cut off from advanced lithography, China has to rely on unproven architectures that struggle with heat and yield. Nvidia hardware beats Huawei’s by about 4.5 to 1 in FLOPS per chip, and Huawei’s parts draw more power for the same performance. That matters as power starts to replace silicon as the binding constraint.
Still, China has plausible routes to large volumes of fairly capable hardware within several years. What matters most is that the US can build even more.
Why compute denial matters
Epoch AI’s October report estimates that AI chips shipped through 2027 could run tens to hundreds of millions of concurrent frontier-model agents. Running nonstop, they could supply as many weekly working hours as 140 to 720 million full-time employees. Blunting China’s access for three to five years or longer would cut the number of agents its companies can run, slowing its AI research and other strategic work.
Can the US slow its own labs?
The analysis splits AI progress into two races.
The first is frontier reasoning, where US labs hold a semi-durable edge from large-scale training, reinforcement learning and test-time compute. Much Chinese progress here appears to come from distilling the reasoning output of US models into systems like Kimi and Qwen. A US lead might allow deliberate slowing, and one popular proposal is a legal cap on compute labs use for research rather than inference. The analyst sees a tension in both restricting Chinese compute and capping American research compute. If efficiency gains move the frontier without more compute, Washington could misjudge its breathing room.
The second race is architecture and efficiency, where the analyst is far less sure of a US lead. Chinese labs including DeepSeek, Moonshot and Qwen keep doing more with less. DeepSeek’s September V4.1-Flash reportedly cut key-value cache memory needs by 75 percent against the prior generation. Cheaper inference means more agents and more experiments on the same hardware.
Some efficiency choices cost monitorability. Moving more reasoning out of explicit chain-of-thought and into latent space is one, and OpenAI drew criticism for trying it in Astra. US industry has largely avoided such looped transformers, but Chinese labs may feel pressure to cut corners, and the analyst sees no easy fix.
A low-confidence forecast and what’s needed
The analyst’s scenario has US hardware controls mostly and temporarily working, US labs voluntarily limiting compute for automated R&D, and Chinese labs still advancing through hard-to-monitor efficiency gains. Meanwhile China’s chipmakers sprint to expand, other countries build multi-trillion-parameter models, and monitorability research improves.
The conclusion is that buying time isn’t winning. Stronger controls are the right policy, but China could catch up faster than many would like. Policy also tends to move after incidents, so the analyst says more research is needed on Chinese fabrication yields, memory output, packaging capacity and how much compute Chinese labs can really reach.
Source: choosingvictory.com