America's AI Chip Controls Are Leaking Faster Than They're Working

In January 2025, a Chinese AI startup called DeepSeek released a reasoning model that matched OpenAI's o1 on key math and coding benchmarks. It did this for a reported $5.58 million in training costs¹, using downgraded H800 chips that exist specifically because of U.S. export controls. That event sent Nvidia's stock tumbling 17% in a single day and launched a debate that still hasn't resolved: are America's chip export controls actually working?

I've spent weeks digging into the evidence, and my answer is uncomfortable for both camps. The controls are doing something. But what they're doing may not be what policymakers think, and the gap between the policy's ambitions and its results is widening at an alarming rate.

Start with the good news, because there genuinely is some. Export controls have made China a marginal producer of advanced AI chips. According to testimony by Commerce Secretary Howard Lutnick², Huawei will produce only about 200,000 AI chips in 2025, compared to the millions Nvidia ships globally. DeepSeek's own researchers found that Huawei's Ascend 910C delivers roughly 60% of the Nvidia H100's inference performance⁵. A Council on Foreign Relations analysis⁶ estimated that Huawei's total AI compute capacity will amount to only 2% of Nvidia's through the second half of the decade. On the hardware production front, the policy is working.

But hardware production and AI capability are increasingly different things. And on capability, the trend is brutal for the containment thesis.

The Stanford HAI AI Index Report 2025³ documented that Chinese models "rapidly closed the quality gap" on benchmarks like MMLU and HumanEval, with performance differences shrinking "from double digits in 2023 to near parity in 2024." The 2026 edition⁴ went further: by March 2026, the gap between the top U.S. model (Anthropic's Claude Opus 4.6) and China's best shrank to just 2.7% on Arena scores, down from a massive lead for GPT-4 in 2023. That's not narrow-benchmark cherry-picking. Arena scores reflect broad capability as rated by human users. The gap hasn't just narrowed. It has nearly vanished.

Now, a defender of the current policy will point out, correctly, that we can't observe the counterfactual. Maybe without export controls, Chinese models would be ahead by now. That's a fair point, and I take it seriously. The Chinchilla scaling laws and everything we know about compute scaling suggest more chips = better models. DeepSeek's own technical innovations, like its Mixture-of-Experts architecture that activates only 37 billion of 671 billion parameters per forward pass¹³, were explicitly developed as workarounds for H800 bandwidth limitations. You don't build expensive workarounds for constraints that don't bind.

So yes: the controls impose real costs. The question is whether those costs translate into a durable strategic advantage. And here's where the evidence turns against the policy.

The leakage problem is not a side issue. It is the central issue. In December 2025, the DOJ announced Operation Gatekeeper⁷, dismantling a network that smuggled $160 million worth of restricted Nvidia H100 and H200 chips⁸ to China through straw purchasers and rebranding schemes. In March 2026, prosecutors charged Super Micro Computer's co-founder with diverting $2.5 billion⁹ in Nvidia-powered servers to China through Southeast Asian front companies. Industry reports cited by Introl¹⁰ estimate at least $1 billion in restricted chips shipped to China in 2025 alone. And in February 2026, a senior Trump administration official told Reuters¹¹ that DeepSeek had trained its latest model on Nvidia's most advanced Blackwell chips, reportedly housed in a data center cluster in Inner Mongolia.

I want to sit with that last point for a moment. Blackwell chips are Nvidia's cutting-edge product. They are explicitly banned from export to China. And yet U.S. intelligence believes DeepSeek is training on them. This isn't a story about old H800s being used creatively. This is about the most restricted hardware ending up exactly where it's not supposed to be.

The pattern here is unmistakable: update controls, circumvent controls, repeat. A Singapore-based company called Megaspeed¹² imported at least $4.6 billion worth of Nvidia hardware and 136,000 GPUs, with more than half being Blackwell chips. When Nvidia inspected the data centers, only a few thousand chips were on-site. The rest had vanished into the supply chain.

This brings me to the deeper structural problem. The AI field is undergoing a paradigm shift from training-time compute to inference-time compute, and the controls were designed for the old paradigm. Researchers at ICLR 2025¹⁶ demonstrated that scaling compute at inference time can be "more effective than scaling parameters" for reasoning tasks. Deloitte projects¹⁴ inference will account for roughly two-thirds of all AI compute in 2026, up from a third in 2023. An Introl analysis¹⁵ projects inference will exceed training compute demand by 118x by 2026.

Why does this matter for export controls? Because inference-time scaling changes the economics of compute advantage. DeepSeek-R1 matched o1 partly by generating 10-100x more tokens per query¹⁵, effectively trading training compute for inference compute. If a 7B parameter model with 100x inference compute can match a 70B model with standard inference, then the training compute gap that export controls protect becomes less decisive. The controls were built to restrict the fuel for a race that's changing engines.

Recent reporting from The Information¹⁷ suggests OpenAI and Anthropic are themselves moving away from pure inference-time reasoning and back toward baking capabilities into training. This could re-validate the training compute thesis. But it also means the frontier is moving in ways that a static control regime cannot track.

Here's where I land. I think the controls-are-useless crowd overstates their case. Controls have genuinely hobbled China's domestic chip production, limited its cloud infrastructure exports, and forced Chinese labs into expensive architectural workarounds. Those are real costs. But the controls-are-working crowd has a worse problem: they cannot point to a durable, measurable capability gap that the policy has maintained over time. Stanford's data shows the gap closing from double digits to 2.7% in three years. The smuggling evidence shows billions in restricted hardware reaching China despite enforcement. And the technological shift toward inference-time scaling is eroding the specific advantage the controls were designed to protect.

The honest assessment is that export controls are necessary but are failing in their current form. Not because the theory is wrong, but because (1) enforcement is catastrophically under-resourced relative to the economic incentives for circumvention, (2) the policy lacks allied coordination with teeth, and (3) the definition of what constitutes a strategic compute advantage hasn't been updated to match how frontier AI is actually being built.

The Trump administration's own moves have made this worse, not better. Rescinding the Biden-era AI Diffusion Rule in May 2025¹⁸ created a regulatory gap that Chinese firms exploited through cloud-based access. The December 2025 decision to allow H200 sales to China¹⁹ while simultaneously prosecuting smugglers of the same chips sent an incoherent signal. And BIS has seen a steady exodus of staff²⁰ since early 2025, weakening the very enforcement capacity the policy depends on.

What should readers watch? Three indicators will tell us whether this policy trajectory is sustainable. First, watch the Stanford HAI gap metric: if it compresses below 2% by the 2027 report, the capability argument for controls will be effectively dead regardless of the hardware gap. Second, watch whether BIS staffing and enforcement budgets increase; if they don't, the smuggling problem will only scale. Third, watch whether the next major Chinese frontier model claims to have been trained on Huawei hardware rather than smuggled Nvidia chips. If it does, and the performance is competitive, it will mean China's domestic chip ecosystem has crossed a threshold that makes the entire control regime structurally obsolete. I think there's better than even odds we see at least two of these three within eighteen months.