A November 13, 2026 Reuters investigation identifies more than $3 billion in Chinese artificial-intelligence computing projects built around restricted Nvidia processors, even though Nvidia has said U.S. export controls effectively shut it out of China's AI data-center market. The story follows Jensen Huang's November 12 resignation and lands the same week Chinese labs reportedly raced to lock overseas GPU leases (our November 12 coverage).
Reuters is not recycling one headline number. It read procurement text and shows who plans which chips in which provinces. That is actionable for compliance engineers and for builders wondering whether remote-access crackdowns will bite harder than on-paper export bans.
TL;DR
| Claim | Source detail |
|---|---|
| Total identified | More than $3 billion in project value |
| Evidence base | Tenders, a government prospectus, company disclosures, university contracts |
| Named silicon | H100, H200, H800 among restricted generations |
| Geography | Coastal Fujian to Inner Mongolia, plus universities |
| Delivery status | Mixed — some operating clusters, many plans without proof of chip arrival |
| Nvidia China AI data-center market | Nvidia has said export controls effectively shut it out; the projects still name its chips |
Relay outlets including the Economic Times and Business Times Singapore republished Reuters' project list with the same caution: planned value is not delivered silicon.
What Reuters documented project by project
The following examples come from Reuters reporting as relayed by Economic Times on November 13, 2026. Dollar conversions use the yuan figures Reuters printed.
Fujian — Xianyou county. A proposed 6.2 billion yuan (~$873 million) AI computing center seeks commercial partners and names Nvidia H100 chips for large-scale AI development.
Shaanxi. State-owned Sinomach said a subsidiary won a contract as part of a project for a 3 billion yuan computing center specifying H100 and H200 processors.
Gansu. A local company awarded a 1.16 billion yuan construction contract in July for a computing center planning about 3,072 Nvidia processors, per tenders specifying H100 and H200 hardware. The contract went to a consortium led by a subsidiary of state-owned Shaanxi Construction Engineering. The documents do not establish whether the processors were delivered.
Jilin. An April government prospectus describes a 2.26 billion yuan project mixing H100 with Chinese-developed accelerators, with about 80% bank financing.
Inner Mongolia. A 1.5 billion yuan plan proposes 625 H200 servers; industry reporting said construction began on two buildings in a first phase.
Operating or disclosed clusters. Reuters also points to operational or announced sites:
- Jiangsu / Wuxi — municipal government said in March 2025 that a cluster using H800 entered operation.
- Fujian — Shanghai-listed Damingcheng disclosed in June an H800 cluster; its computing-services agreement, originally worth 759 million yuan, ended early.
- Universities — Shenzhen University bought three H100 units in September; Northwestern Polytechnical University (U.S. blacklisted) bought two H200 chips in October, per separate Reuters reporting cited in the relay.
Each line is a compliance and logistics puzzle: some are mega-facilities, others are three-GPU teaching labs. Together they show Nvidia naming conventions survive deep into provincial budgets, not only ByteDance-scale orders we covered in Reuters neocloud demand pieces.
What the story does not prove
Reuters is careful, and explainx.ai echoes that care:
- Procurement language is not possession. A tender mentioning H100 does not mean customs cleared a Blackwell shipment yesterday.
- Project value is not chip spend. Yuan figures cover whole projects; they are not price × GPU count.
- Grey-market routing stays opaque. The piece complements, but does not replace, investigations into Singapore routing and entity-list procurement.
Still, the map matters politically. Congress already tied Huang's resignation to export politics in our Speaker Johnson consensus post and Wyden–Neal investigation write-up.
Why this is not the same as "China labs rent Singapore GPUs"
Yesterday's Reuters exclusive described Chinese AI firms reserving overseas Nvidia capacity because they fear Remote Access Security Act momentum after the Huang scandal. That story is about jurisdiction hopping.
Today's review is about domestic paperwork that still names forbidden SKUs. If Washington closes the remote GPU loophole (Commerce rule reporting), projects with onshore H800 clusters become legacy risk, while greenfield tenders may pivot to Huawei Ascend or hybrid Chinese chips like Jilin's prospectus already sketches.
Builders training inside China should plan for three clocks:
- Contract clock — what the tender promised.
- Enforcement clock — BIS, DOJ, and FINRA-style probes after Megaspeed (timeline post).
- Product clock — when open Chinese weights catch up enough to justify redesign.
Export and shipping stack symbolizing restricted Nvidia AI chips moving through global supply chains
Smaller purchases still name Nvidia in tenders
Reuters' relay highlights university-scale buys, not only billion-yuan parks. Shenzhen University awarded a contract for three H100 processors in September. Northwestern Polytechnical University, which appears on U.S. export control lists, bought two H200 chips in October according to separate Reuters reporting cited in the Economic Times piece.
Those orders matter for compliance training: a procurement officer copying boilerplate from a 2021 tender template can still publish forbidden SKUs in 2026 PDFs. Enforcement teams do not only chase hyperscaler leaks; they read campus IT budgets.
Our Northwestern H200 post walked one entity-list case in depth. Today's map shows that story is one dot on a wider scatter plot.
Reading Reuters against Nvidia's own statements
Nvidia has said U.S. export controls effectively shut it out of China's AI data-center market. The November 13 review does not claim Nvidia booked that revenue back. It claims Chinese customers still write Nvidia into plans.
Both can be true:
- Revenue accounting reflects what cleared export law and shipped through sanctioned channels.
- Procurement speech reflects what local officials still want on paper, including via grey routes or delayed deliveries.
Builders should not treat Nvidia's SEC language as proof that no Chinese H800 cluster exists. Treat operational disclosures like Damingcheng's H800 cluster as live infrastructure risk, not historical curiosity.
Policy context after Huang's exit
Reuters' November 12 explainer, summarized by Investing.com and Asia Times, argues tighter curbs are more likely now that Huang — who publicly favored easing China sales — is gone. None of that text, read through accessible relays, announces a signed rule on November 13.
The practical read for developers:
- Expect more audits of who rents whose GPUs, not a sudden open flood of H200 into Beijing racks.
- Watch Southeast Asia datacenter reporting requirements cited in The Information rule summaries.
- Assume procurement PDFs like those Reuters read become exhibits in future enforcement, similar to how U.S. cases used mislabeled server manifests in our Taiwan smuggling coverage.
What builders outside China should do
If you sell compute or models into Asia-Pacific
- Log end-user nationality and IP egress the way AI Overwatch Act proponents want banks to log GPU hours.
- Do not treat "Singapore flag" as sinless; Reuters' domestic map shows Chinese institutions still document Nvidia intent even when renting abroad.
If you benchmark models
- Supply shocks from enforcement can change which API is cheapest mid-project, the same lesson as Sam Altman neocloud double-count claims.
- When a Chinese lab publishes leaderboard jumps, ask which cluster geography hosted the run.
If you fine-tune open weights
- Checkpoint forensics posts like Chinese open weights and Blackwell traces matter more when provincial H100 tenders stay public.
Scenario table for the next ninety days
| Scenario | Likely builder impact |
|---|---|
| Status quo enforcement | Overseas neocloud leases stay attractive; domestic tenders stay on paper |
| Remote-access rule tightens | Chinese labs lose legal paths to Singapore/Malaysia GPUs; prices spike |
| Domestic redesign mandates | Tenders pivot to Ascend plus local chips; CUDA-only code needs ports |
| Criminal cases from Megaspeed threads | Grey-market supply thins; even plan documents quietly drop Nvidia names |
None of these rows is a prediction. They are planning axes given Reuters' evidence plus active Washington politics in our Congress chip-curb coverage.
Questions investors asked after Huang resigned
Equity analysts covering Nvidia often conflate China revenue lines with China demand lines. Reuters' procurement review is a demand line story. It does not, by itself, forecast next-quarter datacenter prints. It does explain why export-control politics stayed on the front page the morning after Colette Kress became interim CEO, as we summarized in Huang's resignation guide.
Funds that still hold Chinese AI ADRs should ask investees which clusters are contracted, powered, and legally operable. A Fujian slide deck that names H100 is useful diligence even when the chips never ship. Model trainers outside China should log the same question when a partner offers cheap API credits backed by unnamed hardware geography. Ambiguity is a compliance cost even when inference still works today worldwide.
How explainx.ai will update this thread
We will extend this post when Reuters publishes delivery confirmations, when BIS posts a matched rule, or when a provincial tender is withdrawn after central review. Until then, treat yuan totals as documented intent, not installed TFLOPS.
Project values and yuan figures follow Reuters as relayed by Economic Times on November 13, 2026; exchange rates fluctuate.