Washington licensed about 10 billion dollars of H200 chips for China, and almost none shipped

The policy fight in Washington was over whether to let the chips go. It ended on 13 January 2026, when the Commerce Department's Bureau of Industry and Security revised its licence review policy so that applications to export the Nvidia H200, the AMD MI325X and chips of similar specification to China would be assessed case by case rather than presumptively denied. The change followed an announcement by President Trump on 8 December 2025. Three conditions attach: the exports must not reduce the semiconductor production capacity available to United States customers, the Chinese purchaser must adopt export compliance procedures including customer screening, and the product must pass independent third party testing in the United States to verify performance and security. Jeffrey Kessler, the Under Secretary for Industry and Security, said in the announcement that permitting the sale of the H200 to China under controlled conditions would strengthen the American technology ecosystem.

Six months later Kessler told the House Foreign Affairs Committee, on 14 July 2026, that the chips had not gone anywhere much. Tech Times, reporting the hearing, quotes him saying that very few shipments against licences for H200s and equivalents have taken place, and that it was a very small quantity of chips, so it is trivial. That testimony was the first official confirmation that any H200 had shipped at all. The licences themselves are estimated at around 10 billion dollars. Roughly ten Chinese firms have been cleared, including Alibaba, Tencent, ByteDance and JD.com, each capped at 75,000 units, buying through Nvidia or through the authorised distributors Lenovo and Foxconn.

The Brookings Institution had already written the obituary. In a commentary published on 17 June 2026, Mark MacCarthy noted that as of mid May 2026 not a single H200 had been sold to a Chinese company, and set that against the 12 to 15 billion dollars Nvidia earned from H20 sales to Chinese customers in 2024 and the 5.5 billion dollar write off it took in April 2025 when Commerce declared the H20 non compliant. His argument is that American leverage over Chinese AI compute has already lapsed, because DeepSeek's models, Huawei's Ascend line and SMIC's manufacturing have made the dependency marginal rather than structural, and because offshore compute in third countries remains available. That is a contested reading and it is his.

The reasons for the gap are not one thing, and the sources disagree about which matters most. Beijing has pushed in the opposite direction to Washington. WinBuzzer reported on 19 May 2026 that no deliveries had started, and that Chinese buyers had domestic alternatives from Huawei, Cambricon, Biren and Moore Threads while they waited for cleaner import terms. Tech Times reports that Chinese regulators have confined H200 use to model training and directed firms towards domestic processors for inference, that officials have told companies to buy foreign chips only where necessary, and that one proposal would bundle each H200 purchase with a set ratio of Chinese made chips. Tech Journal reports Beijing has raised security concerns about the mandatory routing of chips through the United States for certification, viewing it as a possible surveillance channel, and cites Bernstein estimates putting Nvidia's share of the Chinese AI chip market at around 8 per cent against roughly 50 per cent for Huawei.

There is a commercial explanation too, and it cuts against the tidy geopolitical story. Chinese technology companies had collectively ordered more than two million H200 chips for 2026, against Nvidia inventory of roughly 700,000 units, which forced discussions with TSMC about restarting Hopper generation manufacturing. The H200 uses TSMC's 4 nanometre process and CoWoS advanced packaging, the same bottleneck that constrains the Blackwell line. Even a willing buyer would be queuing.

The money the United States takes from the trade is also disputed in its description. Tech Journal, in a piece updated 31 July 2026, describes the arrangement announced in January 2026 not as a tariff but as a revenue share, under which the government receives 25 per cent of the sale revenue of chips routed through the United States for mandatory testing and certification. The same outlet's headline and web address call that arrangement a 25 per cent tariff. The Commerce announcement of 13 January mentions neither a tariff nor a revenue share, and Kessler's testimony did not address it.

One detail from the same hearing has been little noticed. On 14 July, the day Kessler testified about oversight of chip flows, Commerce added ZTE Kangxun Telecom, a unit of ZTE Corp, to the approved buyer list. ZTE Corp was designated a national security threat to federal agencies under the 2018 National Defense Authorization Act.

What is not known is how much of the shortfall is Beijing's choice and how much is Nvidia's supply. Kessler declined to say which licence applications have been denied, though BIS has given Congress a confidential list. Until the split is public, the claim that Washington has lost its leverage and the claim that the chips are simply not built yet both fit the same evidence.