Washington's AI Chip Squeeze Has a Huge Leak
The US approved nearly 880,000 advanced AI chips for China since December despite tightening export controls. Some recipients sit on military watchlists, and new technical analysis suggests Blackwell GPUs may already be training top Chinese AI models.
The Paradox in Plain Sight
The United States has been telling a consistent story for years: advanced AI chips must not reach China. The Biden and Trump administrations alike have tightened export controls, blacklisted companies, and tried to build a wall around the most powerful semiconductors. But newly disclosed congressional data reveals a crack in that wall far larger than anyone expected.
Between December 2024 and May 11 this year, American companies filed approximately 3.4 million applications to ship high-performance AI chips to Chinese buyers. Nearly 879,000 were approved. That is not a rounding error. It is a floodgate.
Who Got the Chips
The approved shipments went to some of China’s most visible technology names. Alibaba received authorization for roughly 75,000 units. Tencent and its subsidiaries, about 150,000. Baidu, roughly 43,000. Electric-vehicle maker BYD was approved for about 4,000. The chips involved include Nvidia’s H200 GPU and AMD’s MI325X — the very products the Pentagon spends billions trying to deny Beijing.
Even more striking, some recipients are not ordinary commercial companies. Four thousand eight hundred H200 chips were licensed for sale by Nvidia to a firm connected to Megspeed, a company currently under US investigation for allegedly smuggling advanced semiconductors into China through back channels. The mere fact that an application was approved — before any determination of wrongdoing — raises urgent questions about how carefully these licenses are being vetted.
Some approved firms also lack clear operational footprints. The New York Times noted that the identity and actual business activities of certain license holders remain uncertain, adding another layer of opacity to a system designed to close off exactly this kind of access.
Blackwell Already Training Chinese AI Models?
Perhaps the most unsettling development came not from Washington but from a technical investigation published in Scientific American. Will Depuy, a former OpenAI researcher, released an analysis tracing a specific hardware signature inside six Chinese AI models: DeepSeek’s V4 series, Moonshot AI’s Kimi K3, MiniMax’s M3, and Ant Group’s Ling 3.1.
Depuy’s method was indirect but ingenious. When AI models train, they initialize with pseudo-random weight values — and the pattern of those initial numbers leaves a faint fingerprint shaped by the underlying GPU architecture. By comparing approximately six trillion possible early-training configurations against known processor layouts, he found matches consistent with Nvidia’s Blackwell GPUs. Controlled experiments reproducing parts of two models confirmed the same result.
Blackwell is restricted from export to China. If accurate, this would mean the most advanced American chips are already powering the next generation of Chinese AI systems — regardless of what the licensing bureaucracy approved or denied.
Depuy cautioned his work remains unpublished and un-peer-reviewed. It does not prove the chips physically crossed the border, only that their computational signature appears in finished model weights. The hardware could have arrived via overseas data centers, rental arrangements, or legitimate re-entry after export — each path existing in legal gray zones.
What This Actually Means
Two things become clear at once. First, the official narrative of tightening strangulation does not match the data. Second, Chinese companies may not need formal access to American chips at all if they can replicate results through indirect routes, third-party jurisdictions, or cloud infrastructure outside US control.
The US government itself acknowledges this tension. Export approval is not the only pathway. Foreign data centers hosting American chips, authorized resellers operating in neutral countries, and companies embedded in complex supply chains all create multiple vectors through which advanced compute reaches China — often faster than regulatory review can intercept them.
There is also a political calculation hidden in plain sight. American semiconductor companies — led by Nvidia and AMD — depend heavily on Chinese revenue. Restricting sales outright would cede market share to competitors elsewhere. Issuing licenses lets Washington maintain leverage without fully severing ties. But every approved shipment erodes the strategic logic behind the controls in the first place.
Who Wins and Who Loses
China wins. Its AI firms gain access to training infrastructure that rivals what domestic manufacturers cannot yet produce at scale. DeepSeek and other Chinese models continue improving even as US policy aims to slow them. Western observers who assumed chokepoint policies would reliably constrain Chinese progress now have reason to doubt that assumption.
American semiconductor firms also win, at least in the short term. Each approved shipment represents revenue that would otherwise disappear if Beijing bought nothing from anyone. Their lobbying influence helps ensure those approvals flow, creating a feedback loop between commercial interest and security policy.
The US national-security establishment loses ground. Every chip that clears export control is one less demonstration that the containment strategy works. The Pentagon’s own military-end-user list should flag dangerous applicants, yet it has not prevented approved shipments to related parties like Megspeed.
Allies caught in the middle face a credibility problem. Japan, the Netherlands, and South Korea have aligned with US export restrictions. But when the United States itself authorizes near-a-million advanced chips for a single rival over five months, the moral authority behind coalition coordination weakens considerably.
What Happens Next
Several trajectories are plausible. One is tightening — Congress pressures BIS to raise the bar on approval standards, require end-use verification, or delay licenses longer. Another is further erosion — administrative inertia and industry lobbying gradually expand the funnel until “controlled” exports approximate unrestricted trade.
A third possibility, perhaps the most likely, is adaptation on both sides. Chinese firms will keep finding indirect paths to advanced compute. American policymakers will keep adjusting the rules, probably with mixed effectiveness. The real question is whether Washington treats this as a temporary leak to be patched or a structural feature of an interdependent economy it refuses to fully decouple from.
If the current trajectory holds, the answer becomes obvious within years rather than decades. Control regimes work best when enforcement matches rhetoric. So far, the evidence suggests the opposite is true.
The Bottom Line
Eight hundred seventy-nine thousand licensed AI chips, four thousand eight hundred routed through an investigated firm, and a former OpenAI researcher claiming to find Blackwell fingerprints in Chinese model weights. These are not anomalies. They are data points — and together they describe a policy environment where the walls are more paper than steel.
The debate over whether sanctions are working has historically been abstract. The congressional documents and technical analysis now make it concrete. The bottleneck is leaking, and nobody in power seems to have closed it.