Alibaba’s AI Compute Has Two Gatekeepers

On Sept. 27, 2026, The Information reported that Beijing had asked Alibaba, ByteDance, and other companies to disclose plans to buy Nvidia RTX Pro 5500 chips, even as officials signaled approval. Washington already reviews advanced-accelerator exports case by case. Alibaba held 23% of China’s AI-cloud market in the first half of 2025, so a purchase decision that once belonged to suppliers and buyers now needs two governments’ consent.

The workstation-class RTX Pro 5500 carries 84GB of GDDR7 memory, compared with 32GB on the RTX 5090 despite otherwise comparable specifications. Beijing’s request reached below the flagship data-center line and into hardware that developers can use for smaller clusters and high-memory AI work.

The Information’s report does not establish a formal procurement rule. Across the 21 months from January 2025 through September 2026, however, Chinese platforms tested domestic chips, rented foreign capacity, and waited for government approvals. Each platform chose its own hedge as both governments changed which hardware could be bought, where it could run, and on what terms.

A purchase order now needs two sovereign yeses

The U.S. Commerce Department placed H200 and AMD MI325X exports to China under case-by-case licensing review. Beijing then added a filter on the buyer side. In January 2026, Chinese authorities reportedly told some technology companies that they would approve H200 purchases only under special circumstances, including university research. Later that month, Beijing reportedly gave DeepSeek conditional approval while regulators finalized the terms.

Each government can stop, resize, or condition the same order for a different reason. Washington applies export policy to the seller and destination. Beijing can apply industrial policy to the buyer, workload, and desired balance between imported and domestic accelerators. A procurement team may have a budget, supplier interest, and an identified workload without an executable transaction.

Beijing has not shut Nvidia out. Chinese authorities reportedly approved a first H200 import batch covering more than 400,000 chips for ByteDance, Alibaba, and Tencent. A first batch that large shows how managed export controls can permit major transactions while still determining who receives scarce foreign capacity.

A100- and H100-based services still rented for less in China than in the United States despite export restrictions. Regulators set the permitted pool, and vendors and customers bargain inside it. Permission now defines the boundary within which the market clears.

Alibaba’s roadmap makes chip continuity release continuity

Alibaba announced an AI strategy spanning models, chips, and data centers on Sept. 22. The company also said it plans to train a model with 5 trillion to 10 trillion parameters. That target does not reveal a training budget, cluster size, or completion date, but it places a large planned model beside an existing cloud business and a widening consumer-AI distribution system.

IDC gave Alibaba 23% of China’s AI-cloud market in the first half of 2025, versus nearly 13% for ByteDance. Alibaba also linked Qwen to Taobao, Alipay, Fliggy, and Amap while targeting a one-stop AI app for 100 million users. Its planned cloud regions in Turkey, Finland, and the Netherlands would add three infrastructure locations if Alibaba completes them within the announced 12-month window.

Alibaba asks the same accelerator supply to do four jobs. Researchers need long training runs, while open-weight releases need evaluation and serving capacity. Consumer services create variable inference loads across shopping, payments, travel, and mapping. Cloud customers expect capacity even when Alibaba’s own products need it. When chips run short, Alibaba’s schedulers must choose whether a scarce cluster trains Qwen, serves Taobao, or remains available to paying cloud customers.

Alibaba, ByteDance, and Tencent buy different kinds of assurance

ByteDance responded to Alibaba’s cloud lead by hiring additional salespeople and lowering prices. The company also reportedly planned to direct roughly 60% of its 2025 AI-chip spending to Chinese suppliers, including Huawei. ByteDance can challenge Alibaba on cloud price while reducing the share of its accelerator budget exposed to imported supply.

Tencent chose another hedge. It backed Enflame, whose Shanghai debut raised about $910 million and produced a $26.3 billion market capitalization after the shares rose 188%. Tencent’s investment gives it exposure to a domestic supplier instead of relying solely on purchase negotiations with established chip vendors.

Beijing has not disclosed an allocation formula favoring any platform by market share, product type, or strategic importance. Alibaba’s 23% cloud share therefore does not guarantee priority. The three companies have built different cases for supply assurance: Alibaba through cloud scale and broad product distribution, ByteDance through supplier diversification and foreign rentals, and Tencent through investment in a domestic chipmaker.

They consequently face different capacity ceilings. Approved imports, domestic accelerators, supplier stakes, and foreign cloud contracts determine how much purchased capacity becomes usable.

Domestic accelerators have become a placement pool

Alibaba, Tencent, Baidu, and other Chinese companies were already testing domestic alternatives by May 2025 as their Nvidia stockpiles dwindled under tighter U.S. restrictions. IDC later estimated that Chinese GPU and AI-chip makers captured nearly 41% of China’s AI-server market in 2025, while Nvidia retained 55%, or roughly 2.2 million cards shipped.

Operators must therefore build around heterogeneous AI compute: selectively approved Nvidia systems where performance and software advantages matter, and domestic systems where supply assurance and policy alignment carry more weight.

Chinese companies have reported that Huawei Ascend chips lag Nvidia in training performance, inter-chip connectivity, stability, and software. An available chip can still require extra engineering, lower cluster utilization, or limit the workloads that teams can place on it.

Huawei and DeepSeek pointed to scale as the answer to uneven substitution. Huawei said its Ascend 950 supernode would fully support DeepSeek V4. DeepSeek separately reportedly planned to deploy more than 160,000 Ascend 950DT chips in an Inner Mongolia data center. The project remains a reported plan rather than a completed cluster, but DeepSeek’s plan treats domestic supply as a fleet to be engineered rather than a one-for-one Nvidia imitation.

Alibaba’s infrastructure teams must match models and services to pools with different memory, software, networking, availability, and approval constraints.

The cluster absorbs the policy shock that the card cannot

Alibaba cannot monetize an accelerator in isolation. Large AI services require clusters spanning hundreds or thousands of servers, along with storage, networking, power distribution, and cooling. The cluster is the operating unit because every weak component can reduce the useful output of the chips around it.

Alibaba can use its cloud infrastructure to route scarce capacity, schedule workloads across clusters, expose inventory to customers, and keep expensive systems utilized. Only after engineers configure and schedule a cluster can procured hardware become a product.

Chinese AI companies reportedly sought compute rentals in Southeast Asia and the Middle East to gain access to Nvidia’s Rubin lineup. ByteDance went further: sources and Nscale’s U.S. filings indicate that ByteDance generated nearly 75% of Nscale’s 2025 sales and used the provider’s Norway facility to access Nvidia AI chips. ByteDance converted the problem of buying restricted hardware into purchasing service from a foreign capacity operator.

Export rules still govern foreign rentals. Even so, buyers without reliable direct ownership favor providers that can assemble chips, power, networking, and legal access in an acceptable jurisdiction.

Alibaba’s planned regions in Turkey, Finland, and the Netherlands are not documented here as an export-control workaround. If built, those regions would give Alibaba more locations in which to assemble permitted infrastructure and serve customers. Geography becomes useful only when Alibaba can combine it with hardware, software, power, and an approved workload.

The Sept. 27 request concerned a workstation-class chip, but it exposed a cloud-scale problem. Alibaba’s 23% share leaves it with more demand to serve while both gates still apply. Its control plane must place each workload on hardware with the right memory, software, networking, geography, and approval. The scheduler turns two governments’ consent into sellable compute.