The United States has approved roughly $10 billion in export licenses for Nvidia's H200 AI chips destined for China, yet nearly all of those chips remain undelivered. A Commerce Department official told Congress on July 14 that while H200 shipments have technically begun, the volumes are negligible. The admission exposed a widening gap between what Washington approves on paper and what actually reaches Chinese technology companies in practice.

The disconnect matters because it reveals how export controls on advanced semiconductors, intended to protect U.S. national security interests, are struggling to keep pace with the realities of global chip manufacturing and distribution. Chinese firms have collectively ordered more than two million H200 chips for 2026, but Nvidia holds only about 700,000 units in inventory. That supply shortage alone explains part of the delivery gap, but regulatory confusion and enforcement gaps tell a more troubling story.

Why Is There Such a Massive Gap Between Approved Exports and Actual Deliveries?

Jeffrey Kessler, the Under Secretary of Commerce for Industry and Security, delivered the blunt assessment to the House Foreign Affairs Committee. "For the American people, the bottom line is very few shipments against licenses for H200s and equivalents have taken place," Kessler stated. "Very small quantity of chips, so it's trivial." The admission marked the first official confirmation that any H200s have shipped to China at all.

"For the American people, the bottom line is very few shipments against licenses for H200s and equivalents have taken place. Very small quantity of chips, so it's trivial," said Jeffrey Kessler.

Jeffrey Kessler, Under Secretary of Commerce for Industry and Security

The Trump administration began issuing conditional clearances for H200 sales in December 2025, approving roughly ten Chinese firms to purchase the chips. Approved buyers include major technology companies like Alibaba, Tencent, ByteDance, and JD.com. Each approved buyer may purchase up to 75,000 chips through Nvidia directly or through authorized distributors Lenovo and Foxconn.

But the licensing machinery has moved faster than actual chip production and delivery. The supply-side problem is real: the H200 uses Taiwan Semiconductor Manufacturing Company's (TSMC) 4-nanometer process and advanced CoWoS packaging technology, the same packaging that has created bottlenecks across Nvidia's entire Hopper and Blackwell product lines. Even willing buyers face manufacturing constraints that regulation alone does not explain.

How Did Advanced Blackwell Chips Reach China Without Detection?

The H200 delivery gap, significant as it is, was not the most alarming revelation from the congressional hearing. That distinction belonged to evidence of a separate, more serious enforcement failure involving Nvidia's more advanced Blackwell chip line. A loophole in Commerce Department enforcement allowed advanced Blackwell chips to flow to Chinese artificial intelligence (AI) firms for nearly a year without triggering export controls.

The mechanism was straightforward but effective. Nvidia's Blackwell, Rubin, and AMD's MI350x-class chips were subject to export license requirements for shipments into mainland China. However, the requirement was triggered by the shipping destination, not by the ultimate owner's headquarters. Subsidiaries of Chinese AI companies incorporated in Malaysia, Singapore, or the United Arab Emirates could purchase these chips without triggering controls, because they were technically outside mainland China. Industry sources cited by Reuters estimated that hundreds of thousands of advanced processors may have flowed through this gap over roughly the past year.

The Commerce Department created the opening in May 2025, when it announced it would not enforce the AI Diffusion Rule issued in the final days of the Biden administration. That rule would have governed global access to AI chips. The Bureau of Industry and Security (BIS) did not close the loophole until May 31, 2026, effectively acknowledging that a rule technically in force since November 2023 had gone unenforced for nearly three years.

Steps to Understanding the Export Control Enforcement Challenge

• Licensing vs. Delivery: The U.S. approved roughly $10 billion in H200 exports, but Chinese firms ordered more than two million chips while Nvidia held only 700,000 units in inventory, creating a fundamental supply mismatch that regulation cannot solve.
• Subsidiary Loopholes: Chinese companies incorporated subsidiaries in third countries like Malaysia and Singapore to purchase advanced Blackwell chips without triggering export controls, a gap BIS acknowledged closing only in May 2026 after roughly a year of exploitation.
• Enforcement Delays: The AI Diffusion Rule existed since November 2023 but went unenforced for nearly three years, allowing hundreds of thousands of advanced processors to potentially reach Chinese AI firms undetected.
• Retroactive Compliance: The May 31 guidance closing the loophole does not require data centers outside China to shut down Blackwell-equipped servers already in service or stop providing computing services on those machines.

The Commerce Department's response to the loophole has been criticized as inadequate. When Representative Bill Huizenga asked Kessler why the May 31 guidance stated that Chinese companies could retain any Blackwell chips they had obtained through smuggling or other loopholes, Kessler responded that companies that had not been licensed for the chips should voluntarily self-disclose the violation. Huizenga's reaction made clear that voluntary compliance was unlikely to be effective.

The same day Kessler testified, the Commerce Department added new names to its list of approved Chinese buyers. ZTE Kangxun Telecom, a unit of ZTE Corp, received authorization to purchase H200 chips. This detail warrants scrutiny because ZTE Corp was designated by the U.S. government as a national security threat to federal executive agencies under the 2018 National Defense Authorization Act, a designation that remains in force. A server manufacturer called Maginfra and a cloud computing subsidiary of Kingsoft also received approvals that day.

The approvals broaden the set of authorized buyers beyond the major internet platforms and large electronics distributors cleared earlier, extending into telecommunications equipment, server manufacturing, and enterprise cloud services. This expansion is occurring while more than 100 Chinese entities approved for Commerce Department blacklisting by an interagency committee last year, including the AI firm DeepSeek and memory chipmaker ChangXin Memory Technologies, remain unpublished on the Entity List, the longest pause in more than a decade.

Beijing's own posture has added friction of a different kind. Chinese regulators have restricted H200 use strictly to AI model training, while directing companies to use domestic processors, primarily Huawei's Ascend line, for routine inference workloads. Chinese officials have separately told domestic technology companies they should purchase foreign chips only when necessary, and one proposal would require each H200 purchase to be bundled with a set ratio of Chinese-made chips.

The situation reflects a broader challenge facing U.S. export control policy: the gap between regulatory intent and enforcement capacity continues to widen as global supply chains become more complex and as companies find creative ways to circumvent restrictions. Whether the Commerce Department can close remaining loopholes and enforce the rules it has established remains an open question.