China's Semiconductor Gamble, One Year Later

In the spring of 2026, the Biden administration’s final wave of semiconductor export restrictions landed with unusual precision. The rules tightened controls on Nvidia’s A800 and H800 chips — the slightly neutered versions Beijing had been buying after the original H100 ban — and extended restrictions to a broader range of chip-design software, advanced packaging equipment, and the high-bandwidth memory that modern AI accelerators require. The stated goal was unambiguous: prevent China from training frontier AI models competitive with American capabilities.

One year on, the honest assessment is that the restrictions achieved some of what they promised and almost none of what their most optimistic proponents predicted. China did not collapse. It also did not catch up. What happened instead was more interesting and less comfortable for both sides than either narrative allows.

Start with what China actually built. Huawei’s Ascend 910C, which began shipping in meaningful volumes in late 2026, is a genuinely capable AI training accelerator. Independent benchmarks — some conducted by researchers at ETH Zurich, some leaked from Chinese state enterprises — suggest it performs at roughly 60 to 70 percent of an H100’s throughput on standard transformer training workloads. That is not a number to dismiss. It is also not parity, and the gap widens significantly when you move from raw arithmetic to the memory bandwidth and interconnect efficiency that large model training actually depends on.

The more revealing data point is yield. TSMC and Samsung achieve yields — the percentage of chips that come off the line working correctly — above 85 percent for advanced nodes. SMIC, China’s most capable domestic foundry, has reportedly been achieving yields in the 30 to 40 percent range for its most advanced 7-nanometer processes. At that yield rate, every chip you actually use costs you two or three chips’ worth of silicon and fab time. The economics are brutal, and they don’t show up in benchmark comparisons.

Baidu, Alibaba, and ByteDance have all deployed Ascend 910C clusters for inference workloads (running models, not training them) where the yield-adjusted cost matters less than raw availability. For inference, the economics are more forgiving. But training the next generation of foundation models — the GPT-4 moment that China wants to own domestically — still requires the kind of cluster scale that domestic hardware makes expensive enough to change strategic decisions.

The domestic LLM story is similarly complex. Baidu’s ERNIE 5, released in January 2027, posted scores on the MMLU benchmark that are competitive with GPT-4 class models. Alibaba’s Qwen 3 series demonstrated strong performance on Chinese-language tasks and respectable performance on English ones. DeepSeek, the research lab backed by the quantitative hedge fund High-Flyer Capital, released a model in early 2027 that the international AI research community treated as a genuine surprise — not because it was ahead of American capabilities, but because it demonstrated that Chinese researchers had internalized transformer optimization techniques that the Western research community had assumed were not yet widely distributed.

The honest question is whether benchmark performance translates to the kind of sustained capability improvement that frontier AI competition actually requires. Training GPT-4 took roughly 25,000 A100-equivalent GPU-days. Training GPT-5 class models, by most estimates, required three to five times that. The next generation requires more again. At Ascend 910C efficiencies and SMIC yields, that arithmetic becomes a serious constraint on how fast Chinese labs can iterate.

There is a secondary problem that gets less attention: the training data pipeline. Frontier model quality depends not just on compute but on data curation infrastructure, reinforcement learning from human feedback at scale, and the red-teaming ecosystem that identifies and corrects model failures. American labs have had years longer to build these pipelines, and many of the researchers who built them are at OpenAI, Anthropic, and Google. The export controls did not restrict the movement of people — but the movement of people, in both directions, has slowed considerably in the current geopolitical climate.

The third-country effects are where the story gets genuinely complicated. The Netherlands, which hosts ASML, implemented its own export controls on EUV lithography machines in early 2023 and has progressively tightened them. The Dutch government did not want to be in this position — ASML is a national champion, a rare case of a European company holding an essentially uncontestable global monopoly, and restricting its sales to China costs it billions in revenue. But Washington applied sustained pressure, and the Dutch eventually aligned.

What the Dutch learned in this process is that they had geopolitical leverage they never asked for and still are not entirely comfortable using. The country’s foreign minister described it in a parliamentary hearing in late 2026 as “an asymmetric responsibility” — the obligation that comes with controlling a chokepoint in a critical supply chain. The phrase is careful, almost diplomatic. The underlying reality is starker: a country of 17 million people, without a permanent seat on the UN Security Council, with no independent nuclear deterrent, holds a card that both superpowers want to play.

South Korea’s situation is different in character but similar in discomfort. Samsung and SK Hynix collectively supply a large portion of the high-bandwidth memory that AI chips require. The US pushed Korea to restrict that memory to Chinese buyers. Korea complied, partially, with significant delays, because its relationship with China is not the same as Germany’s relationship with Russia — China is South Korea’s largest trading partner by a wide margin, and Seoul has legitimate reasons to treat a rupture in that relationship as an economic catastrophe rather than a moral victory.

Taiwan’s position remains structurally unchanged from where it was two years ago, which is to say uniquely dangerous. TSMC continues to operate as the world’s most critical piece of industrial infrastructure, continues to be located ninety miles off the Chinese coast, and continues to navigate the US-China relationship by performing simultaneous loyalty to both sides with a discipline that would exhaust a professional diplomat. The company announced expanded facilities in Arizona, Kumamoto, and Dresden in 2026 — partly genuine strategic diversification, partly insurance, partly an attempt to make itself too embedded in allied economies to abandon.

The most important question the export controls raised — and have not yet answered — is whether they achieved their stated goal. The stated goal was not to prevent China from having AI. It was to maintain a capability gap sufficient to preserve American strategic advantage for long enough to establish irreversible leads in the applications and institutions that matter.

By that standard, the jury is genuinely out. The capability gap in training frontier models is real and probably widening slightly, because American labs can iterate faster on better hardware. The capability gap in deploying AI for specific high-value applications — surveillance, logistics optimization, autonomous vehicle systems — is narrower than the benchmark comparisons suggest, because those applications don’t require frontier model performance. They require good-enough models, deployed at scale, with the kind of patience and integration effort that state-directed industrial programs excel at.

The most unsettling finding from the past year is not that China is ahead or behind. It is that the competition has produced a kind of parallel ecosystem development that the original export control designers did not fully anticipate. China is not trying to build an H100. It is trying to build a system — hardware, software, data infrastructure, institutional knowledge — that doesn’t require an H100. That is a different problem than catching up to a specific benchmark. It is the problem of whether independence itself, at some tolerable performance level, is a viable strategic alternative to parity.

The answer to that question will matter more than any individual chip benchmark for the next decade of AI geopolitics.

One year is not enough time to render a final verdict. It is enough time to notice that the competition is more interesting than the framings suggest, that the third countries are more important than the bilateral framing implies, and that the assumption of a clean American victory — enforced by export controls and sustained by talent advantages — is not yet supported by what has actually happened.

China has adapted. Adaptation is not the same as success. But it is not the same as failure, either.