NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Reuters counted $3.2 billion in China computing projects designed around chips Nvidia says it can no longer sell there

A Reuters review of 150-plus Chinese procurement documents found AI data-centre projects worth over 23 billion yuan (about US$3.2 billion) specified for restricted H100, H200 and H800 chips — including two H200s bought by a blacklisted university for $123,000.

2026-11-18 · 1161 words · NeuroAI
Reuters counted $3.2 billion in China computing projects designed around chips Nvidia says it can no longer sell there

The official story is that Nvidia has left China's AI data-centre market. Export controls shut the door, the company stopped booking Chinese data-centre revenue in its forecasts, and Jensen Huang said last year that its market share there had fallen from 95% to zero.

Then Reuters went through the paperwork. Its 13 November review of more than 150 Chinese government procurement notices, investment plans, construction contracts and corporate disclosures found a paper trail running in the opposite direction.

Key takeaways

  • Reuters identified AI computing projects and equipment agreements with a combined reported value above 23 billion yuan (about US$3.2 billion), all designed around restricted Nvidia H100, H200 or H800 processors (Reuters, 13 November 2026, as carried by NetEase and other outlets).
  • The total covers whole projects — buildings, power systems, supporting infrastructure — not chip purchases, and spans planned, contracted, under-construction and partially operating facilities.
  • Examples range from a proposed 6.2 billion yuan (about US$873 million) computing centre in Fujian planned around the H100, to a Gansu contract specifying roughly 3,072 Nvidia processors whose arrival remains unconfirmed.
  • Northwestern Polytechnical University, on the US Entity List since 2001, bought two H200 chips on 21 October for 874,700 yuan (about US$123,000) from a domestic supplier, per a procurement record on its own website.
  • Research firm IDC estimated Nvidia still accounted for 34% of China's AI accelerator shipments in the first half of 2026 — a different measure, but one that also sits awkwardly next to "zero".

What the 150 documents show

The projects Reuters found sit across coastal tech hubs and western provinces — government-backed data centres, university computing facilities and enterprise installations. A few examples:

  • Fujian: a planned AI computing centre in Xianyou county with a stated value of 6.2 billion yuan (about US$873 million), specified for H100 chips and still seeking commercial partners.
  • Shaanxi: a 3 billion yuan (about US$420 million) computing centre designed for H100 and H200 chips; a subsidiary of state-owned engineering group Sinomach disclosed that it had won a construction contract on the project.
  • Jilin: a 2.26 billion yuan (about US$318 million) project planned to mix H100 chips with domestically made processors, with developers expecting roughly 80% of funding to come from bank loans.
  • Inner Mongolia: a 1.5 billion yuan (about US$211 million) plan specifying 625 H200 servers; local industry reports say the first two buildings broke ground this year.
  • Gansu: a data-centre retrofit that awarded a 1.16 billion yuan (about US$163 million) construction contract in July, with tender documents calling for roughly 3,072 Nvidia processors including H100s and H200s. Reuters noted the documents do not confirm whether the chips have actually arrived.
  • Jiangsu: a municipal H800 computing cluster in Wuxi has been in service since a March 2025 city announcement, and Shanghai-listed developer Damingcheng disclosed this June that a computing venture had built an H800 cluster under a 759 million yuan (about US$107 million) services agreement that was later terminated early.

Reuters was explicit about the caveat: 23 billion yuan (US$3.2 billion) is the value of projects and equipment agreements, not what China has spent on Nvidia chips. Planned deployments are not delivered deployments, and none of the documents alone shows a breach of US export rules.

The blacklisted university that bought two H200s

The sharpest single data point comes from a procurement record, not a plan. On 21 October, Northwestern Polytechnical University (西北工业大学) published a purchase record on its website showing it had bought two H200 chips for 874,700 yuan (about US$123,000) for its management school. The supplier was Shaanxi Bosuan Network Technology, a domestic firm — not Nvidia directly.

That matters because the Xi'an university has been on the US Commerce Department's Entity List since 2001, over alleged missile-proliferation ties, with restrictions tightened in 2019. It is one of the "Seven Sons of National Defense" (国防七子) universities, and licences for restricted exports to it are normally denied.

The university's procurement records also show five contracts this year renting access to Nvidia-powered servers through Chinese companies. Its cybersecurity school signed two seven-month contracts in January — worth 441,700 yuan and 441,000 yuan (about US$62,200 and US$62,100) — for servers holding eight H200 chips each, for model fine-tuning and reinforcement-learning training. Reuters reviewed 26 such records and found the same pattern elsewhere: Shenzhen University bought three H100s for 769,800 yuan (about US$108,000) in September, and Beihang University paid 489,000 yuan (about US$69,000) for four A800s in April.

In January, the US administration began reviewing H200 export applications for approved Chinese customers case by case. That change did not touch the separate restrictions on Entity List organisations. Chips may also be reaching buyers through resale or rental rather than direct sale — a gap between policing who buys chips and policing who uses the compute.

Two numbers that measure different things

The awkward juxtaposition: Nvidia says export controls have shut it out of China's data-centre market and excludes China data-centre revenue from guidance, while IDC estimated it took 34% of Chinese AI accelerator shipments in the first half of 2026.

The two figures can both be true: Huang's "95% to zero" describes Nvidia's legal, direct sales channel, while the IDC figure counts hardware physically arriving in China, possibly not sold by Nvidia itself. Reuters claimed no more, and no public finding has linked these projects to violations of US law. But the gap explains why US lawmakers keep asking who ends up using the compute, not just who signs the purchase order.

Honest limitations

The 23 billion yuan (US$3.2 billion) total aggregates documents of different kinds — tenders, plans, contracts, disclosures — at different stages, some of which may never be built out. It includes data halls, power systems and other infrastructure, so it overstates chip spending by design. Most documents do not show whether specified chips were delivered, through which channel, or whether they remain in service. The 34% IDC figure measures shipments, not sanctioned sales, and Reuters itself could not determine how the restricted chips in its sample entered China. No identified project has been tied to violations of export law, and Nvidia, the universities named and the US Commerce Department's Bureau of Industry and Security did not respond to Reuters' requests for comment. Treat this as a map of intent and paperwork, not an audit of silicon.

What readers can do now

  • If you track export-control effectiveness, watch the two ends Reuters flagged: resale intermediaries and rental compute contracts, both of which let end users reach restricted chips without appearing on a customs form.
  • If you run a data-centre project in China designed around H100/H200-class hardware, assume the supply question is unresolved, and price the cost of a domestic-accelerator fallback — Reuters noted that switching architectures mid-project means reworking servers, software and cluster design.
  • If you invest in the sector, note that Nvidia's guidance already excludes China data-centre revenue, so any licensed resumption would be upside to guidance — while enforcement actions against intermediaries are a separate risk channel the documents now make visible.

Related coverage

Cite this article

NeuroAI (脑机网), "Reuters counted $3.2 billion in China computing projects designed around chips Nvidia says it can no longer sell there," 18 November 2026, https://neuroai.site/a/na-corp-nvidia-china-projects-230b

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Companies & Stack” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories