China’s $3 billion Nvidia project pipeline faces a delivery test
Editorial Team
••📖6 min read
Chinese computing projects still specify Nvidia processors, but their reported values mix proposals and contracts. Conditional US licenses, Chinese restrictions, financing and customer demand complicate the path to usable capacity.
A review identifies more than $3 billion in Chinese computing projects built around restricted Nvidia processors, spanning proposals and operating clusters.
The figures mix project budgets, construction awards and services contracts; they do not measure Nvidia chip revenue.
US licensing, Chinese restrictions, financing and customer demand each present a separate test of commercial viability.
More than $3 billion in Chinese computing projects remain tied to Nvidia processors despite export restrictions, according to a November 13 project review. The evidence ranges from facilities seeking investors to clusters reported to be operating. That makes the pipeline consequential, but uneven: a proposed investment, a construction award and a working computing business represent different commitments.
Nvidia, the US supplier of graphics processing units and accompanying AI software, says US controls have effectively excluded it from China’s AI data-center market, the review notes. A quarterly filing qualifies that picture: some licensed shipments have occurred, while Chinese restrictions have also constrained sales.
The review appeared a day after Jensen Huang, Nvidia’s co-founder and an advocate of easing export restrictions, resigned as chief executive amid a scandal. Nvidia said it appointed Colette Kress as interim CEO; she had been its chief financial officer and executive vice president since 2013. The leadership transition comes as developers still need to turn specifications for Nvidia hardware into financed, authorized and commercially useful capacity.
H100 SXM accelerator in Nvidia’s March 2022 launch materials. Image: Nvidia. Source: NVIDIA.
One pipeline, several stages of commitment
The projects specify an established computing platform. Nvidia introduced the H100 GPU in March 2022 as its first processor based on the Hopper architecture, succeeding Ampere. Its announcement described hardware and software tools for applications including large AI language models. Those capabilities explain what developers are specifying; they do not establish why each chose Nvidia over another supplier.
Treasury Secretary Scott Bessent said agencies had backed export restrictions on Nvidia customer Megaspeed for months. He challenged Commerce’s handling of the case and Nvidia’s safeguards, while Megaspeed denies illicit chip exports to China.
Editorial Team
The project review identifies five large developments with materially different scopes:
Location
Reported amount
What the review establishes
Xianyou county, Fujian
6.2 billion yuan ($873 million)
Proposed AI computing center seeking commercial partners and specifying H100 chips
Shaanxi
3 billion yuan
Computing-center project specifying H100 and H200 processors; state-owned engineering group Sinomach said a subsidiary won a contract within it
Jilin
2.26 billion yuan
Planned project combining H100 processors and Chinese-developed chips
Inner Mongolia
1.5 billion yuan
Development proposing 625 H200 servers; industry reporting in June said construction had begun on two first-phase buildings
Gansu
1.16 billion yuan
Construction contract awarded in July for a center planning approximately 3,072 Nvidia processors, with H100 and H200 hardware specified
The five entries are examples, not a reconciliation of the review’s headline total. Their amounts cannot be treated as equivalent chip orders: a whole-center proposal and a construction contract cover different things. Nor does a server count establish a processor count.
For Gansu, the review says tender documents identify the intended hardware but do not establish delivery. The construction consortium was led by a subsidiary of state-owned Shaanxi Construction Engineering. A building award demonstrates a different kind of progress from receiving and commissioning processors.
Financing is another unresolved stage. The Jilin developer expects bank loans to supply about 80% of its investment—roughly 1.8 billion yuan of the proposed 2.26 billion. That is an expectation of borrowing, not evidence that lenders have disbursed it.
Jilin’s inclusion of Chinese-developed chips also limits the idea of an exclusively Nvidia-dependent pipeline. Domestic alternatives are part of at least one plan. The review supplies no comparable prices or performance measurements that would establish the cost of substituting them.
Company-reported H200 figures: quarterly revenue share, an inventory and purchase-obligations charge for the first half of fiscal 2027, and the U.S. import tariff. These are supplier-level disclosures, not project costs. Source: NVIDIA — quarterly filing for July 26, 2026.
H200 access is conditional, with barriers on both sides
The US Commerce Department’s Bureau of Industry and Security, which administers these export controls, revised its October 2022 advanced-computing restrictions in October 2023. The agency said the restrictions were intended to limit technology supporting China’s military modernization. That history does not provide a complete account of the rules governing later purchases.
A rule effective January 15, 2026 changed US license review for H200 chips and qualifying equivalents or less advanced chips from a presumption of denial to case-by-case consideration. Conditions include US supply protections, recipient security procedures and independent testing in the United States. Conditional review is not permission for every buyer or project.
Nvidia’s quarterly filing for the period ended July 26, 2026 adds a commercial counterweight. The company said Chinese government restrictions constrained licensed H200 sales. Shipments under the US licensing program represented less than 1% of its Data Center revenue in that quarter.
Nvidia also reported a $400 million H200 charge for excess inventory and purchase obligations in the first half of fiscal 2027 as demand diminished. It said the program’s US inspection route subjected shipments to a 25% tariff upon importation into the United States, which it had been unable to pass on to customers.
Those are supplier-level disclosures, not losses or costs assigned to the projects in the review. They nevertheless show why US permission alone does not settle demand, delivery or the economics of a sale. Neither a listed chip model nor a project announcement establishes that a particular transaction was authorized—or that it violated a restriction.
Operating hardware does not guarantee lasting revenue
Some projects have moved beyond plans. The review cites a March 2025 announcement by Wuxi’s municipal government that an H800 computing cluster had entered operation. It also cites Shanghai-listed Damingcheng’s June disclosure that a computing venture in Fujian had built an H800 cluster.
Damingcheng’s customer contract supplies a separate warning about commercial durability. The review says a computing-services agreement originally worth 759 million yuan ended early. That figure describes services for a customer, rather than a payment to Nvidia, and installation did not ensure the agreement ran its course.
Demand also extends beyond large developers. The review cites separate reporting that Shenzhen University awarded a contract for three H100 processors in September, while US-blacklisted Northwestern Polytechnical University bought two H200 chips in October. Those small procurements broaden the picture of demand without establishing delivery across the much larger project pipeline.
The next test is project-level evidence
For developments still seeking money or hardware, the next evidence needed is committed financing, transaction-specific authorization, documented deliveries and commissioning. The review does not supply comparable processor acquisition costs, electricity expenses or service prices from which to judge returns across the projects.
For clusters already installed, the test shifts to utilization and durable paying customers. Damingcheng’s early-ended agreement makes that distinction material: functioning hardware and sustained commercial demand require separate evidence.
Applied Materials expects chipmaking equipment demand to rise more than 30% in 2027, while supplier capacity and factory readiness will determine how much demand becomes sales.
Shi Bingzhe and two companies face money laundering charges over 917 servers allegedly acquired from Aperia International. The case now covers five individuals and six companies, but the new charges do not establish the servers’ chip contents or ultimate destination.
Nothing’s December AI device has appeared with red and green lights and an apparent recording button. The clues suggest an audio tool, but existing AI recorders leave the company needing to explain its hardware, service costs and use of personal context.
Apify says an intruder accessed customer accounts and personal information or stored data. Customer notices describe credential rotations and two-factor resets, while the affected-account count and full data exposure remain unclear.
The former Huawei server business has passed Shenzhen's listing committee review for an 8 billion yuan IPO. Its growing market share and profits sit alongside negative operating cash flow and a sharp increase in short-term borrowing.
EU leaders gather for their first dedicated AI discussion, weighing investment, foreign dependencies and safety while Ukraine diplomacy competes for attention. The informal summit is a strategic discussion, with no political statement scheduled for adoption.
New hardware orders have increased Nscale's 2027 funding needs and limited its ability to defer spending. The British AI infrastructure provider is pursuing a New York listing targeting a $35 billion valuation while financing for its Anthropic contracts remains uncommitted.
Commerce Secretary Howard Lutnick has agreed to voluntary Senate testimony as bipartisan scrutiny of Nvidia exports turns to government approvals and enforcement. A hearing date remains unconfirmed, and intelligence suspicions remain disputed.
The startup founded by Microsoft data-system veterans has raised $70 million across seed and Series A rounds. Its unified analytics platform targets unpredictable AI-agent workloads, but the claimed savings still need comparable performance evidence.
Samsung Electronics is reportedly in talks to buy Modine as it expands a cooling business already bolstered by FläktGroup. Modine brings fast-growing data center sales, but weaker margins and a Texas factory still ramping production complicate the potential deal.
Nvidia faced pressure to enlarge its financing backstop before Jensen Huang departed. The leadership change raises questions about support for AI customers, but interim CEO Colette Kress helped build that strategy and chip demand remains a counterweight.
Google launched Gemini 4 Hydrogen and automatic model switching in Antigravity, claiming near-Carbon task success with less quota and faster completion. Lower token rates support the efficiency pitch, but reasoning settings, mixed competitive results and inconsistent benchmark labels leave the cost of completed work to be tested.
Japan is considering safety and reporting duties backed by enforceable orders for advanced AI. The proposal would strengthen its nonbinding framework, but capability thresholds, covered businesses and penalties remain unresolved.
Google says its new fast AI model is reaching Search users worldwide, improving complex answers and source discovery. The rollout expands access to Gemini 4, but its effect on answer quality and website traffic remains unmeasured.
Four nonprofits are asking a California court to stop five AI developers from building or deploying more capable systems. The case seeks enforceable limits even as OpenAI has previously reported a training pause and is launching defensive AI tools.
Expo Simulators gives coding agents remote iOS and Android environments with recordings for reviewers. The launch extends Expo's cloud development services into verification, with per-minute charges and shared capacity limits.
Kleiner Perkins and Clearlake Capital led Mirendil’s $1 billion Series A. Its agent-led model-building experiment points to smaller research teams, but company benchmarks, substantial compute costs and a private beta qualify the promise of wider access.
Mount Thor raised $45 million in equity, alongside separate hardware financing, to expand Apple infrastructure for AI agents and model inference. Existing Mac clouds and workload-specific performance limits put its demand and cost claims to the test.
Salesforce Ventures led a Series C that values Q-CTRL above $1 billion. The funding backs GPS-free navigation and quantum computing software, while undisclosed revenue and contested performance benchmarks leave the commercial case open.
