The U.S. Approved 750,000 Nvidia Chips for China. None Arrived. Here's What Did Instead.

Here is one of the stranger facts of the four-year-old U.S.–China chip war: the United States authorized the sale of three-quarters of a million of Nvidia's most advanced Artificial Intelligence (AI) chips to Chinese companies — and not a single confirmed unit was delivered.

That is the headline finding of "Covert Compute," a new investigation by C4ADS, a Washington-based nonprofit research organization largely funded by the U.S. government, published in early September and widely covered this week.

The number is 750,000 Nvidia H200 accelerators — approved for roughly ten Chinese firms, each capped at 75,000 units, with Foxconn licensed as the distributor. As of mid-2026, according to the report's analysis of trade and customs data, actual deliveries stood at zero.

At the same time, C4ADS traced controlled Nvidia chips entering China through three documented "gray" channels — including one opaque corporate network it valued at $4.6 billion.

If you want to understand the actual state of the AI chip standoff, the story is not "banned or not banned." It is something much messier: legal chips that never move, and restricted chips that move anyway — while China builds its own supply in plain sight.

Where the 750,000 Number Comes From

To read the number correctly, you need the mechanism behind it.

Export controls on AI chips have escalated in stages since October 2022, when the U.S. first restricted Nvidia A100 and H100 sales to China. Each Chinese workaround chip — the A800, the H800 — was eventually closed off. In April 2025, even the deliberately downgraded H20 was put under license requirements, forcing Nvidia to take multibillion-dollar inventory charges.

Then policy reversed. In December 2025, the U.S. said it would allow H200 sales to China under a revenue-sharing arrangement. On January 15, 2026, the Bureau of Industry and Security (BIS) shifted the H200 and AMD's MI325X from a "presumption of denial" to case-by-case review, attached to a 25% levy on transaction value paid to the U.S. Treasury.

Around ten Chinese companies — reports name Alibaba, Tencent, ByteDance, and JD.com among them — received approval, with a 75,000-unit ceiling each. Foxconn was authorized as distributor. The theoretical total: 750,000 chips worth, analysts estimate, on the order of $20–25 billion at market prices.

The licenses never became shipments. U.S. Commerce Secretary Howard Lutnick acknowledged the standoff in congressional testimony: Chinese buyers, he said, were holding back to concentrate investment in their domestic industry. Nvidia's Chief Financial Officer (CFO) confirmed on an earnings call that the company had booked zero revenue from H200 sales in China. Chinese customs had paused H200 imports within hours of the January authorization, according to multiple reports.

In other words, for once both sides agree on the facts — while attributing them to opposite motives. Washington says Beijing is protecting Huawei. Beijing does not much dispute it.

The Three Gray Channels

While the legal channel sat frozen, C4ADS documented — using Chinese government filings, Asian trade statistics, and corporate records — three routes by which controlled Nvidia hardware still reached Chinese buyers between 2022 and January 2026:

The details on routes 1 and 2 are granular. Chips shipped from Taiwan to Vietnam — where testing and packaging facilities give transshipment plausible cover — then onward to Hong Kong, a customs territory Beijing treats as "outside" the mainland for trade purposes. Two Hong Kong entities, Profit New Limited and ELB International Limited, dominate the import records; Profit New alone declared $8.7 million in silicon imports in a single day in March 2025, using paperwork C4ADS calls deficient by U.S. regulatory standards but that Chinese authorities did not challenge.

The Megaspeed case is by far the largest. The company was the biggest Nvidia hardware importer in Southeast Asia from 2023–2025, owned until 2023 by Chinese firm 7Road Holdings, then nominally by Singapore-registered Swiftdata — with corporate records pointing to a Chinese businesswoman, Huang Le, as a principal. Beneficial ownership of the parent remains untraced. C4ADS notes Megaspeed appears to have sourced Blackwell-generation chips, whose export to China is flatly prohibited.

How much does this matter in volume terms? Epoch AI, an AI research evaluator, has estimated that anywhere from one-third to a majority of China's AI compute may run on Graphics Processing Units (GPUs) obtained outside official channels. The estimate is broad for a reason: gray-market flows are, by construction, unmeasurable precisely.

The Fine Print the Report Itself Insists On

Responsible coverage of "Covert Compute" requires pausing on the caveats, because C4ADS builds several of them into its own methodology:

First, customs records are not convictions. The report states explicitly that import declarations and procurement documents do not, by themselves, prove that transactions were completed or that laws were broken. Routes are described in the language of indicators — ownership opacity, paperwork irregularities, implausible end-user profiles — not proven smuggling verdicts.

Second, the totals are floors, not ceilings. C4ADS only counts transactions where specific chip models were explicitly named. Shipments mislabeled as generic electronics would not appear. The real gray volume, the authors repeatedly note, is likely higher than documented.

Third — and most important for the $4.6 billion figure — C4ADS does not claim all of Megaspeed's hardware went to China. It documents the volume, the ownership trail, and the risk indicators; Southeast Asia itself has legitimate, fast-growing AI demand. Conflating the entire network with covert Chinese imports would overstate the report's own conclusion.

Fourth, consider the source's position. C4ADS is a respected open-source-intelligence organization, but it is funded principally by the U.S. government, and its recommendations — more BIS inspectors, on-the-ground end-user audits abroad, post-delivery verification — are policy advocacy aimed squarely at tightening the existing control regime. That does not undermine its data work; it is simply where the organization sits.

What China Bought Instead

The backdrop to zero H200 deliveries is that China's buyers were not idle — they were redirecting the money.

C4ADS and industry trackers point to the scale-up of Huawei's Ascend line in 2026:

• Ascend 910C production reportedly doubled to roughly 600,000 units in 2026
• Total Ascend-series die output across all models projected at 1.6 million units
• Tencent, DeepSeek, and Moonshot AI are all scaling domestic-hardware deployments
• ByteDance placed an order reported around $5.6 billion for Ascend capacity earlier in the year; its trillion-parameter LongCat 2.0 was trained on Ascend

International Data Corporation (IDC) data from June 2026 had domestic chips at 41% of China's 4 million AI accelerator cards shipped in 2025, with Huawei the largest domestic vendor. Analysts at Bernstein estimate Huawei took roughly half of China's AI chip market in 2026, while Nvidia's share — about 95% two years ago — has fallen to single digits. Nvidia chief executive Jensen Huang has described his company's China data-center position as effectively zero, and Nvidia no longer models Chinese recovery into its long-range forecasts.

This is the strategic logic of Beijing's choice to leave 750,000 approved chips on the table. Buying the H200 would have injected a competitor's strongest hardware into the market at the exact moment its own ecosystem crossed from porting models to natively training them — in exchange for a 25% payment to the U.S. Treasury. From Beijing's seat, the expensive chips were the cheaper option not to take.

The Paradox at the Heart of the Controls

Step back, and "Covert Compute" captures a genuine paradox that neither side's political framing resolves.

The formal control regime has, by its own narrow metric, worked: the most advanced chips do not flow legally at scale, and the U.S.–China model-capability gap has held broadly stable at roughly 6–12 months, per independent analysts. U.S. policy was designed to slow frontier AI development with military and intelligence applications, and on that score researchers call it "broadly successful."

But three leaks run underneath that success: gray-market hardware that keeps arriving, cloud access that lets Chinese users rent foreign compute remotely, and model distillation that extracts capability from exported Application Programming Interface (API) access. Epoch AI's one-third estimate is the quantified version of the leak.

Meanwhile, the controls produced an outcome Washington did not design for: a Chinese customer base that, having been shown it can be cut off at any moment, now treats full-stack independence — chips, interconnect, software, models — as non-negotiable. The H200 licenses were supposed to be a release valve. China welded the valve shut itself.

What to Watch Next

Three signals will tell whether the 750,000-zero gap is a moment or a new equilibrium:

• Does any H200 ever ship under the license? The framework remains legally live. A first confirmed large delivery would signal Beijing has decided domestic supply alone is not enough; another year of zero would confirm the divorce.
• Do gray-channel estimates harden or fade? If independent researchers keep finding one-third-plus of Chinese compute on undocumented hardware, the control regime's enforcement gap becomes its defining feature — and the likely target of new multilateral rules.
• Does Ascend's 2026 scale-up hold? Huawei's 1.6-million-die target is, like all fab plans, hostage to yields and manufacturing inputs. Hitting it converts "zero H200 deliveries" from a sacrifice into a non-event.

For outside observers, the lesson of the C4ADS report is that the chip war stopped being a simple blockade story. The legal channel carries nothing. The illegal channel carries an unknown but substantial amount. And the domestic channel — the one that will define the next decade — is being built fastest of all.

The U.S. offered China 750,000 of the world's best AI chips. China said no. That single refusal may be the most revealing transaction in the entire standoff — even though it never happened.