What Reuters actually counted
Start with what the $3.2 billion is not. It is not $3.2 billion spent on Nvidia chips. Reuters' total is the reported value of whole projects: data-center buildings, power systems and construction contracts, as well as the processors named inside them. It also mixes three different stages of reality, from an investment plan seeking partners, to a signed construction contract, to one cluster in Wuxi that Reuters reports has been running H800 chips since March 2025. ALFW reviewed DailyMech's and Seoul Economic Daily's accounts of Reuters' reporting, not Reuters' original or the Chinese documents.
Read that way, the number is still useful. Procurement notices and tender documents are public paperwork. Across more than 150 of them, Chinese local governments, state contractors and universities are still writing Nvidia's restricted parts into their plans: H100 servers in Fujian, H100 and H200 in Shaanxi and Gansu, H100 paired with domestic chips in Jilin, 625 H200 servers proposed in Inner Mongolia. The documents record proposed or contracted specifications; they do not establish eventual procurement.
Demand is not delivery
The review does not close the gap it opens. Reuters says plainly that the documents do not establish how restricted hardware reached Chinese users, and a tender naming 3,072 processors is not evidence that 3,072 processors were delivered. Some of these projects may end up running domestic accelerators, fewer Nvidia chips than planned, or nothing at all. The Xianyou center, the largest item in the tally at 6.2 billion yuan, is still looking for commercial partners.
That distinction separates this story from the forensic work published earlier this week, which found Blackwell starting-weight fingerprints in six Chinese open-weight models. That was evidence about hardware actually used to produce those weights. Reuters' tally is evidence about hardware wanted, at project scale, in the open, in government paperwork.
The awkward middle for Nvidia
Nvidia told investors in August that export controls had effectively shut it out of China's data-center market. An IDC estimate cited by Reuters puts Nvidia at 34% of Chinese AI-accelerator shipments in the first half of 2026 anyway. The figures concern different periods and measures: Nvidia's August statement is about market access, and IDC's is a first-half shipment estimate. They do not establish how the chips reached buyers.
The timing gives the tally weight beyond its caveats. It landed a day after Jensen Huang's resignation on November 12, and days before Nvidia reports results on Tuesday, November 17, under interim chief executive Colette Kress. Any update on China revenue or guidance will now be read against projects with reported values totaling about $3.2 billion, including infrastructure and plans not yet realized.
Sources
- [1] Reuters, review of Chinese AI projects built around restricted Nvidia chips, published November 13, 2026 — read via full-text relays: DailyMech (read Reuters' exclusive and companion factbox in full); Seoul Economic Daily: https://en.sedaily.com/world/2026/11/14/chinas-nvidia-powered-ai-projects-top-32-billion-reuters ; AI Agent Update summary: https://aiagentupdate.com/post/china-ai-projects-restricted-nvidia-chips-2026-11-14Read source
- [2] ALFW prior coverage for context (not a source for the new facts): Blackwell fingerprint piece, November 13, 2026Read source
- [3] ALFW prior coverage for context (not a source for the new facts): Nvidia China share piece, October 13, 2026Read source