Chinese AI Data Centers Still Specify Restricted Nvidia Chips

A Reuters review of more than 150 Chinese procurement notices, investment plans, construction contracts and corporate disclosures found over 23 billion yuan ($3.2 billion) in computing projects and equipment agreements designed around Nvidia processors restricted by US export controls.

Reuters reported the review on November 13, 2026.

The 23 billion yuan total is not a chip-sales estimate. Reuters said it spans planned investments, signed contracts and completed installations, including buildings, power systems and other infrastructure. The figure therefore measures the broader Chinese computing buildout rather than spending on Nvidia processors alone.

The reviewed projects stretch from coastal technology hubs to remote western provinces. They also extend beyond major technology companies to local government-backed ventures, universities and regional computing centers, according to Reuters.

Nvidia has said US controls effectively shut it out of China’s AI data-center market. The chipmaker also excluded China data-center computing revenue from its forecasts, citing regulatory uncertainty, Reuters reported.

The documents show a different implementation picture. Reuters found Chinese customers continuing to plan purchases and, in some cases, operating restricted processors. The company’s exclusion from direct sales has not removed Nvidia hardware from Chinese project designs.

For infrastructure buyers and Washington, the distinction is consequential. A named processor in a tender is not proof of delivery, and an operating cluster does not identify its end user. Reuters’ review keeps those procurement and end-use questions separate.

Which projects still specify restricted Nvidia chips?

Reuters’ companion project factbox identifies five proposed or contracted developments:

• Xianyou county, Fujian: A 6.2 billion yuan AI computing center seeks commercial partners and specifies H100 chips for large-scale AI development.
• Shaanxi: A Sinomach subsidiary won a contract tied to a 3 billion yuan computing-center project specifying H100 and H200 processors.
• Jilin: A 2.26 billion yuan project combines H100 processors with Chinese-developed chips. Its developer expects bank loans to finance about 80% of the investment.
• Inner Mongolia: A development valued at 1.5 billion yuan proposes 625 H200 servers.
• Gansu: A local company received a 1.16 billion yuan construction contract for a computing center planning approximately 3,072 H100 and H200 processors.

The entries cover different stages of development. Xianyou is seeking commercial partners, the Jilin developer expects bank loans, and the Gansu company received a construction contract. Reuters’ factbox records those arrangements without converting planned processor counts into verified operating capacity.

Reuters said the Gansu tender documents do not establish whether the processors were delivered. That leaves a gap between the facility’s design and its confirmed hardware. A signed contract can settle construction terms while leaving the chip question open.

The pipeline matters to Nvidia’s market position even though the company excludes China data-center computing revenue from its forecasts. Reuters found planned projects across government-backed ventures, universities and regional centers still specify restricted processors. Direct-sales exclusion and continued hardware design are separate conditions.

Which restricted chips are already operating?

Some cases in Reuters’ review move beyond tender language. Wuxi’s municipal government said in March 2025 that an H800 cluster had entered operation in Jiangsu. The official statement establishes operation, but Reuters said the documents do not establish how the restricted hardware reached Chinese users.

Damingcheng disclosed that a Fujian venture had built an H800 cluster and signed a computing-services agreement originally worth 759 million yuan. Damingcheng also disclosed that the contract ended early.

The disclosures weaken any simple claim that export controls have removed Nvidia hardware from China’s operating computing base. They do not establish the route taken by the H800 systems, their ultimate users or the purposes of the capacity. Reuters’ review treats end use as a separate question from the identity of a buyer.

How could export controls change the pipeline?

Washington began restricting Nvidia’s advanced AI processors in 2022, Reuters reported, and later tightened controls on modified chips developed for China, including the H800.

The Trump administration has also moved to review H200 export licenses for eligible Chinese customers individually, subject to conditions. The reviewed projects name H100, H200 and H800, tying three processor lines to the controls Reuters examined.

For facilities still seeking hardware, Reuters identified three risks: delays, supply disruptions and costly redesigns around domestic chips if controls tighten further. Those risks make the pipeline sensitive to policy changes even where a project has already specified an imported processor.

Reuters did not establish that the planned processors in the proposed facilities had been delivered. The review instead separates a processor listed in procurement documents from hardware confirmed at a site. That distinction leaves announced capacity subject to license timing, supply availability and redesign costs.

Washington is therefore trying to separate access from use. The 2022 restrictions and later H800 limits tighten particular routes, while individual H200 license reviews create a conditional channel for eligible Chinese customers. The project records still leave the users of installed capacity unresolved.

What do the documents prove about diversion?

Reuters’ documents do not establish how restricted hardware reached Chinese users. The review also found no link between the projects and Megaspeed, or between the projects and the allegations involving Jensen Huang.

Huang resigned as Nvidia’s chief executive on November 12, 2026, one day before Reuters published the review. Reuters found no public finding that Huang violated export-control laws.

Megaspeed has denied illegally diverting Nvidia chips to China. That denial remains separate from what procurement documents show about planned hardware and reported H800 operation.

The legal record leaves two questions open without resolving the broader enforcement problem. The documents show chip specifications and some operating clusters; they do not establish the ultimate users or purposes of restricted computing capacity.

Sunny Cheung, an analyst at the Washington-based Jamestown Foundation, said the findings reinforce his own view on enforcement: “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.”

Cheung’s focus matches the evidentiary gap Reuters identified. A surface buyer can differ from the organization that ultimately uses the capacity, and the reviewed project set includes local government-backed ventures, universities and regional computing centers.

For policymakers, end-use tracking is therefore a separate requirement from tracking export licenses for future shipments. The stakes are practical: Washington must determine who controls restricted capacity, not only which company appears in a purchase agreement.

Reuters’ records separate two practical tests that remain open. First, announced facilities must show whether they can secure the processors named in their designs. Second, regulators must identify who uses restricted capacity already installed.

The next checkpoint is concrete: whether planned H100 and H200 capacity becomes delivered hardware under current controls, and whether regulators can identify the end users of H800 clusters already operating.