US Approves 900,000 AI Chips to China Despite Military Ties
Washington granted broad AI semiconductor export licenses to Beijing—many going to firms on the Pentagon's military watchlist—sparking fresh doubts over whether national security reviews were ever meaningful.
How Washington Lost Its Edge
The numbers are not large by any traditional trade yardstick. Eight hundred seventy-nine thousand units of advanced computing silicon. But stacked together, they expose something far more consequential than any single customs statistic: the American export-control regime lost its coherence sometime between December 2024 and May 2025, and no one in public view seemed to notice until Congress forced it into the open.
According to documents published by the House Foreign Affairs Committee on November 19, the Commerce Department licensed the export of 879,000 artificial-intelligence chips produced by NVIDIA and AMD to Chinese firms during that five-month window. The figure covers applications—not just confirmations of physical shipment—but the gap between what Washington permitted and what actually crossed the Pacific remains opaque enough to matter. NVIDIA disclosed to Bloomberg that only roughly 10,000 H200-class chips appear to have reached China through legitimate channels; the rest of the approved volume may never have left American ports.
That distinction, however small it sounds, carries enormous weight. It means the Trump administration authorized far more than it ultimately allowed—another way of saying the licensing machinery operated on paper while actual compliance depended on corporate discretion.
But here is where the story thickens beyond routine export-control accounting: several of the approved recipients sit directly on or have recently surfaced on the Pentagon’s blacklist of firms suspected of supporting the People’s Liberation Army.
Tencent was placed on “List 1260H”—the Congressional-mandated registry of Chinese companies linked to the military—in January 2025. Alibaba, Baidu and BYD appeared on that same register in February before being temporarily removed, only to be formally reinstated in June of this year.
During the period under review, those four firms collectively received 272,000 of the approved chips. Tencent alone was allocated 150,000. Alibaba took 75,000. Baidu received 43,000. BYD, the electric-vehicle manufacturer, obtained 4,000.
The type of silicon assigned to each firm was not disclosed publicly. That omission matters because the applications filed with Commerce covered roughly 3.4 million units across both manufacturers—an overshoot of fourfold against what was actually licensed—and included high-performance products such as the H200 GPU and AMD’s MI325X series, architectures far closer in capability to restricted variants than to the downgraded versions originally intended for the Chinese market.
The Security Review Question
Chris McGuire, who served as deputy assistant secretary for technology and national security on the National Security Council under Joe Biden’s White House, called the finding “shocking.” Thirty percent of all chips exported under Commerce authorization went to enterprises on the military enterprise black list, he told Reuters, adding that the episode raised fundamental doubts about whether the department had ever conducted a serious national-security review before approving the applications.
McGuire’s concern echoes a broader unease inside the Beltway that the Trump administration, often criticized abroad for its lax trade posture toward China, has been equally careless on the technological side. The irony is sharp: an executive branch frequently accused of prioritizing commercial deals over strategic restraint ended up permitting massive semiconductor flows that would have faced near-certain denial under the previous administration’s tighter controls.
NVIDIA responded to Bloomberg on the record by insisting that its licenses came with strict conditions and limited volumes, and acknowledged that while certain permits were granted, it has since suspended outbound H200 shipments to China. AMD stated merely that it complies with all applicable U.S. export laws and regulations—a deliberately thin answer for a company whose entire high-end revenue pipeline now runs through Washington’s permission slip.
What neither firm addressed directly is the question of what happens to the chips already authorized but not yet physically shipped, or worse, those that may already be en route through transshipment routes not captured in official tallies.
What Chinese Firms Are Saying
The reactions from China tell their own story.
Alibaba warned investors in its recent earnings call that new U.S. regulatory uncertainty could disrupt parts of its overseas data-center expansion plans—a remarkably blunt acknowledgment that the firm’s infrastructure strategy was built around the very supply lines Washington now appears unwilling to guarantee consistently.
Tencent offered a more general observation: irrespective of whether Chinese companies lease computing resources abroad or build them domestically, the net result is the same. Access to frontier-grade processing power is contracting.
Both statements imply that the firms receiving the chips understood the geopolitical environment imperfectly when they accepted the licenses. They also suggest that even when Commerce grants permission, the surrounding policy atmosphere carries enough volatility to make long-term planning difficult—a paradoxical outcome for any administration claiming that controlled engagement stabilizes relationships.
Who Wins and Who Loses
In the short term, NVIDIA and AMD win nothing in a structural sense. Their shares may tick higher on license announcements, but every approved chip that subsequently lands on a restricted-entity list erodes the credibility of the export-control framework itself. The framework’s credibility is the asset these companies implicitly sell to Washington every time they apply for a permit; once it cracks, future negotiations become harder, not easier.
The Chinese firms gain capacity in the near term—computing power that accelerates domestic AI research and product development. They lose in the medium term because the episode ensures that future restrictions will target them even more tightly. Every approved license today plants the seed for a prohibition tomorrow.
Washington loses the most intangible but arguably the most important prize: the appearance of coherent policy. When the executive branch authorizes flows that lawmakers publicly denounce as inconsistent with national-security interests, the signal to allies—especially Japan and the Netherlands, which coordinate closely with the United States on semiconductor controls—is unmistakable. Recalibration becomes inevitable rather than optional.
The Bigger Picture
This is not a scandal in the conventional sense. No evidence suggests corruption or deliberate bypassing of rules. What it does reveal is an administrative system struggling to keep pace with the speed at which semiconductor demand migrates across borders, especially when the products in question sit squarely at the intersection of commerce, security and geopolitics.
The Trump White House entered office promising deregulation and deal-making flexibility. Those priorities did not vanish overnight—they simply collided with a reality that export controls do not operate by presidential decree alone. They depend on interagency review, statutory requirements, and the willingness of individual firms to self-restrict beyond what the law demands. None of those mechanisms performed coherently during the five-month window.
Congress, through the House Foreign Affairs Committee, exposed the discrepancy. That exposure may prompt legislative tightening, agency recalibration or further corporate self-restraint. It will almost certainly not end the debate about how much computing access China should have while advancing capabilities that blend civilian and military utility.
China’s entry into this particular conversation arrived quietly through earnings calls and investor presentations. The real reckoning, however, will come from answers that have yet to be publicly given: exactly how many of the 879,000 approved units were ever shipped; how many reached their designated recipients; and what role—if any—middlemen and third-country logistics played in closing the gap between Washington’s authorization and the final destination of advanced silicon.