China continues building AI centers around banned Nvidia chips
Two investigations published on November 13, 2026, show that Chinese projects remain designed around Nvidia chips subject to US controls. AI models trained in China would also bear traces of Blackwell.
Despite export controls imposed by Washington, Chinese AI infrastructure projects continue to be designed around restricted Nvidia chips, and some models published in China appear to have been trained on these accelerators. Reuters has identified over 23 billion yuan in projects, contracts, and investments related to this type of hardware, while a former OpenAI engineer says he has found the “signature” of Blackwell in several Chinese models. The issue is no longer just about who buys the chips, but who ultimately uses the computing power.
Des projets encore dessinés autour de GPU interdits
In Reuters’ investigation published on November 13, 2026, more than 150 purchase notices, investment plans, construction contracts, and company documents were scrutinized. The result: over 23 billion yuan, or 3.2 billion dollars, in computing projects designed around Nvidia processors subject to U.S. restrictions. This figure does not correspond to chip purchases alone; it also includes buildings, electrical systems, and associated infrastructure. In other words, the technical specifications sometimes outlast commercial policy.
These projects are not only concentrated in major coastal hubs. They also extend to western provinces and involve local governments, universities, and regional computing centers, in addition to private actors. In several cases, tender documents mention H100 or H200 chips without proving that the processors were actually delivered. This gray area is central: a project can be designed around prohibited hardware without confirmed physical presence of that hardware.
Entre H100, H200 et H800, le casse-tête continue
The examples cited by Reuters illustrate this gap. In Xianyou, Fujian, a 6.2 billion yuan AI computing center is seeking commercial partners and mentions H100 chips for major AI developments. In Shaanxi, a Sinomach subsidiary won a contract related to a 3 billion yuan project specifying H100 and H200 chips. In Jilin, a 2.26 billion yuan project combines H100 chips with domestically developed chips, with expected bank financing covering about 80% of the investment.
Other cases go beyond mere intention. In March 2025, the municipality of Wuxi announced that an H800 cluster had been commissioned in Jiangsu. Damingcheng also stated that a joint venture in Fujian had built an H800 cluster and signed a computing services agreement initially valued at 759 million yuan, although the contract ended prematurely. Again, the supply chain resembles a labyrinth rather than a straight line.
Quand le matériel se cache dans les modèles
The second source, published the same day, presents another signal: Chinese AI models would bear the trace of Nvidia Blackwell chips despite export restrictions. Will DePue, a former OpenAI engineer, claims to have found this footprint in models from DeepSeek, Moonshot AI, MiniMax, and Ant Group. His approach is based on the idea that certain model weight initializations depend on the GPU and seed used initially, leaving an exploitable signature after training. In short, the hardware leaves digital crumbs.
According to DePue, six Chinese models were initialized on Blackwell systems, including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3. Four of these correspond to Grace Blackwell configurations, meaning Nvidia’s GB200 and GB300 components in its liquid-cooled rack systems. However, the author acknowledges a significant limitation: finding a GPU’s footprint does not reveal the physical location of the chips. A Chinese lab could have used this hardware via an American supplier, leaving a loophole that U.S. legislative proposals aim to close [to be verified].
La question n’est plus seulement la vente, mais l’usage final
Both investigations converge on one point: U.S. controls complicate direct sales but are not enough to prevent the circulation of sought-after computing capacity. Nvidia has stated for several months that U.S. restrictions have effectively excluded it from the Chinese AI data center market and has removed this revenue item from its China forecasts. Yet, documents identified by Reuters show that clients continue to plan for these chips in their projects and sometimes already use them.
Sunny Cheung, an analyst at the Jamestown Foundation, sums up the enforcement challenge: “Washington must stop focusing solely on who buys the chips on the surface and start examining who ultimately uses the computing power—and for what purpose.” This is precisely where the next showdown will take place. As long as a project can be designed with a prohibited chip, then financed, resized, or rewritten along the way, the line between compliance and circumvention will remain blurred. And tender documents will continue to pretend not to see anything.
By the numbers
- 23 billion yuan — projects and contracts identified in China by Reuters, November 13, 2026.
- 150 documents — purchase notices, plans, and contracts examined by Reuters.
- 6 models — Chinese models Will DePue says are linked to Blackwell.
- 759 million yuan — initial value of a service agreement in Fujian, according to Damingcheng.
- 80% — expected share of bank financing for a Jilin project, according to Reuters.