business 7 min read

US Approved 880K AI Chips for China. They May Never Arrive

Washington has green-lit 880,000 advanced processors for Beijing-based firms — yet a growing number sit unshipped, revealing a containment strategy that looks stronger on paper than on the dock.

  • NVIDIA
  • US-China Tech War
  • AMD
  • China AI
  • Semiconductor Sanctions

The Paper Win That Isn’t

In the US-China chip war, Washington keeps declaring victory even as its own controls appear to be unraveling.

Last month, details surfaced from a congressional records request showing the US Commerce Department had approved approximately 879,900 semiconductor shipments to Chinese companies — out of 3.4 million total applications. On its face, the math suggests restraint. Roughly one-quarter approval rate means American regulators are still drawing a line somewhere.

But the numbers miss the more interesting story, and arguably the more important one: most of those approved chips have not actually reached their Chinese recipients.

This discrepancy matters because it reveals a crack in America’s containment architecture that goes beyond any single company’s compliance failures. An approved license is not the same as delivered silicon, and that gap tells us more about how the policy works in practice than the raw approval figures ever could.

The Aliquots That Stayed Behind

Alibaba, Tencent, ByteDance, and Baidu each received export license approvals for 75,000, 75,000, 75,000, and 43,000 NVIDIA chips respectively. JD.com and DiDi also held approved allocations. None have apparently taken physical delivery.

AMD told a similar tale. Kuaishou, DiDi, a Tencent subsidiary, and a ByteDance subsidiary each received licenses for 75,000 units — again, no visible shipments following.

Most of these unfulfilled approvals were applications submitted after December 2025, according to media reports citing released government documents. Earlier records have not been made public. Whatever happened before that date remains unclear, and the silence itself is telling.

Chinese companies are getting their names on permits. The permits are being issued. And then nothing follows.

These are not small orders. Each allocation represents hundreds of millions of dollars in potential revenue for American chipmakers and a meaningful performance tier above what domestic Chinese alternatives currently offer. The gap between signed authorization and shipped hardware suggests that the bottleneck sits somewhere between the Bureau of Industrial Security’s desk and the shipping container bay.

What Happens Between Signature and Shipping

Export controls are not just a legal instrument. They’re an enforcement reality.

When the Bureau of Industrial Security grants a license, the next step involves commercial negotiation, logistics, end-use monitoring, and — critically — the willingness of exporters to comply with the spirit of the regulation, not just its letter.

Here, several pressures converge. NVIDIA and AMD operate under enormous political and reputational risk from Washington. Both companies’ executives have testified before Congress multiple times about compliance obligations. For either firm, quietly shipping products to restricted-end customers — even ones holding a valid license — would invite scrutiny far worse than anything that comes from a failed deal. A single violation can trigger criminal referrals, sweeping civil penalties, and debarment from future contracts.

That’s one scenario. Another is simpler: license applicants queue up, file requests hoping for approval, but don’t pressure sellers to deliver because they know the deal might die after the paperwork clears. Companies may apply in large numbers precisely because the process is long and uncertain, banking on eventual clearance rather than committing resources to expedite anything.

Or perhaps the applications themselves were strategic — filing to secure some form of precedent or bargaining position rather than to obtain immediate hardware. In industrial policy, presence on a permitted list carries weight. It signals that a company remains recognized, eligible, and partially integrated into the global technology ecosystem. That signal matters to investors, partners, and competitors regardless of whether a single unit ships.

All of this means an approval count is a poor proxy for how much compute power is actually flowing into China. The metric the world watches — total approved applications — conflates intent with outcome.

The Broader Regulatory Tightening

The White House understands the loophole problem well enough to act beyond traditional border controls.

On November 16, the Commerce Department published new rules targeting remote access to American-made AI infrastructure. Chinese firms attempting to lease computing capacity through data centers in third countries — Malaysia among them — now fall within the regulatory perimeter. Overseas foundry arrangements involving Chinese chip design firms are similarly constrained.

The message is blunt: the US does not trust physical inspection alone. It wants jurisdictional reach that extends across supply chains and across borders. If enforced comprehensively, that expansion raises difficult questions about sovereignty, extraterritoriality, and whether other nations will simply build alternatives outside the US orbit.

But the policy also reveals a more practical reality. Border-level controls alone are insufficient when compute can be provisioned remotely, when design work moves through intermediary jurisdictions, and when procurement can be disguised through complex corporate structures. The new rules attempt to close these pathways by extending US regulatory authority to services rendered abroad using American hardware and software stacks.

If this approach gains traction, it reshapes the competitive landscape for every country seeking AI capability without relying on American infrastructure — including the United States’s closest allies.

The Political Layer

Lawmakers are watching closely, and not all of the attention is benign for the companies involved.

Hunterbrook Capital, a US investment firm, publicly accused AMD last week of maintaining business relationships with Chinese firms subject to export restrictions, triggering a short-seller campaign. AMD shares fell nearly 11 percent at one point and closed down about 7.7 percent. That kind of market punishment reflects how quickly sentiment can shift when Washington turns.

Meanwhile, lawmakers flagged that NVIDIA had applied in December to sell 445,000 H200 chips to Tencent — a volume so large it attracted suspicion about possible use of intermediary channels with uncertain ties. Questions from Capitol Hill rarely stay quiet. Even unsubstantiated allegations can move markets and chill customer confidence.

The Pentagon list of Chinese military-affiliated companies includes Alibaba, Tencent, Baidu, and BYD. That designation doesn’t carry an outright embargo but restricts access to defense contracts and certain financial services, reinforcing the broader chilling environment. Major banks and payment processors have independently tightened scrutiny on transactions involving listed entities, creating a de facto secondary sanction layer that operates without formal legislation.

Taken together, these dynamics mean that compliance is no longer just a legal exercise. It is a boardroom priority, an investor-relations concern, and a factor that shapes hiring, product roadmaps, and market strategy for companies with billions in revenue at stake.

Who Wins, Who Loses, What Comes Next

American policymakers want credit for controlling the flow. Beijing wants evidence that its firms are succeeding despite restrictions. Neither side gets a clean narrative right now.

For Chinese AI developers, the immediate consequence is frustration rather than collapse. The approved licenses represent a tangible signal that doors remain partly open — but only on paper. Actual silicon procurement continues to depend on informal channels, gray-market sourcing, repurposed older-generation GPUs, and increasingly on domestic alternatives like Huawei’s Ascend series.

Those alternatives are improving. Huawei’s latest training clusters now rival earlier-generation NVIDIA offerings on throughput per dollar, though they lag on software ecosystem maturity and interoperability with Western toolchains. Every quarter, the gap narrows slightly. Restrictions that once seemed decisive are losing their sharpness over time.

For NVIDIA and AMD, the uncertainty is costly. Compliance teams must staff complex review processes, manage political risk, and navigate a regulatory environment that shifts faster than product roadmaps. Revenue from China — historically a meaningful share of both companies’ totals — now carries an optionality tax: every sale requires assumption of worst-case enforcement scenarios.

For the US government, the risk is reputational rather than material. Every unshipped approved license weakens the narrative that export controls are working as intended, even if enforcement mechanisms tighten further. The appearance of policy failure can matter more than its substance when it comes to deterrence.

The real question is whether America prefers to be seen restricting flows or actually restricting flows. So far, the records suggest the former wins in committee briefings while the latter lags behind in customs houses. The deeper structural challenge is that containment policies tend to accelerate the very capabilities they seek to suppress — not through malice but through the straightforward logic of incentives. When the easiest path to high-performance compute is blocked, capital and talent reroute toward homegrown alternatives. Each denied shipment is a subsidy to domestic resilience, however unintended.

For now, the numbers tell a paradoxical story: Washington approves hundreds of thousands of chips it cannot reliably prevent from arriving, then tightens rules on every alternate route those chips might take. The strategy is expanding in scope while its evidentiary base grows thinner. That tension is unlikely to resolve soon. It defines the current phase of the technology cold war.