China Requires Domestic AI Chips in Data Centers Receiving State Funding

A New Rule Reaches Into China's AI Construction Boom

China has issued guidance requiring new data center projects that receive state funding to use only domestically produced artificial intelligence chips. Projects less than 30 percent complete have been told to remove foreign accelerators already installed or cancel planned purchases. Facilities further along in construction will be assessed case by case.

As of November 9, 2025, officials have not published the directive, and its geographic reach remains unclear: it may apply nationwide or in selected provinces. The change nevertheless goes beyond earlier policies that discouraged purchases of foreign chips. It makes domestic hardware a requirement for the affected projects and reaches back into procurement decisions already made.

The financial stakes are substantial. Since 2021, AI data center projects in China have attracted more than $100 billion in public funding linked to national and provincial goals for data sovereignty. Those goals concern control over the infrastructure that stores and processes data. Foreign suppliers dominated the first wave of construction, making the new requirement a major change in how publicly supported computing capacity is built.

Nvidia's H20, the most capable AI accelerator the company was permitted to sell in China, falls within the restrictions. The guidance also targets more powerful H200 and B200 processors. US export rules have restricted those chips, but some have continued to reach Chinese buyers through gray market channels, outside normal authorized sales routes.

Domestic suppliers including Huawei, Cambricon, Enflame, MetaX, Moore Threads and Biren are positioned to take the business. For operators, the immediate task is more complicated than choosing a different supplier: replacing accelerators can require changes to servers, cooling, networking and the software used to run AI models.

Which Projects Must Change Their Purchases?

The guidance applies to facilities receiving government money in any form. Support can include grants, tax incentives, discounted land, assistance with electricity prices or direct government equity stakes. A project can therefore fall within the rules even when public support comes through reduced operating costs rather than a payment toward construction.

Operators have been instructed to audit procurement plans and reopen bidding for domestic chips where foreign parts were planned. Integration partners, which assemble components into working computing systems, are updating their bills of materials. These lists identify the parts required for each system and must now reflect the domestic hardware requirement.

Projects in the early stages of construction are the priority. Some have already been paused before breaking ground, including a facility in a northwestern province that planned to deploy Nvidia chips. The requirement to remove installed foreign accelerators also means that an early project cannot simply keep equipment because it has already arrived.

Cybersecurity and economic planning regulators are involved in the reviews. For projects further along, decisions will depend on timing, the computing work the facility needs to perform and the availability of domestic hardware. That leaves advanced projects facing individual assessments while operators of projects below the 30 percent threshold must revise their equipment plans.

Why Replacing Chips Can Mean Rebuilding Systems

Operators ordered to remove foreign accelerators face several stages of work. Teams must take servers out of service, record the equipment being removed, cancel outstanding orders and qualify replacement boards and racks built around domestic chipsets. Qualification means checking that the new components work together reliably and meet the site's requirements.

Power delivery, cooling loops and the networks connecting servers are designed around particular hardware. A different accelerator may draw electricity differently, release a different amount of heat or use another connection system. Vendors will need to recertify racks, and some facilities may have to rebuild aisles.

Facilities close to completion face a different set of possible outcomes. Some may be permitted to keep existing foreign hardware so they can begin service on schedule, then move to domestic chips over time. Others may be required to restrict foreign processors to workloads considered less critical or limit their use.

These possible arrangements would affect how operators divide work between old and new equipment. In all scenarios described by the guidance, operators must demonstrate that new computing capacity will rely on domestic silicon. The practical challenge is to keep a facility usable while changing the hardware foundation on which its expansion depends.

Nvidia Loses Ground as Chinese Suppliers Gain Space

Nvidia, AMD and Intel are the clearest foreign suppliers to lose from the policy. The restriction removes their accelerators from the largest pool of growth in China: the public sector and cloud infrastructure backed by public money.

Nvidia designed the H20 to comply with US export rules, but it is now off limits for the affected state projects. H200 and B200 processors already face US export restrictions and will attract additional scrutiny if they appear in sensitive facilities. Inspectors reviewing procurement records and inventories will make the continued use of chips obtained through gray market channels more difficult.

Nvidia once controlled more than 90 percent of China's accelerator market, with its share at about 95 percent in 2022. Its share has now fallen to zero. The shift closes the door on a return through custom processors designed for the Chinese market and redirects purchasing toward local suppliers.

Huawei's Ascend is the most established domestic line, with devices used for both AI training and inference. Training is the process of building a model by learning patterns from data. Inference is the work a trained model performs when it answers a question, makes a prediction or handles another request.

Cambricon makes machine learning units that fit into standard servers, while Enflame designs accelerators for data centers. MetaX and Moore Threads have introduced graphics processing unit families, and Biren focuses on processors for large training systems. Server manufacturers Inspur and Sugon are supplying systems built around domestic chips. Alibaba Cloud, Tencent Cloud and Baidu AI Cloud have also showcased clusters running on local silicon.

The Software Switch Is as Important as the Hardware

Developers often prefer Nvidia because of CUDA, its software stack for programming accelerated computing. CUDA provides thousands of optimized kernels, the small routines that carry out particular calculations, alongside mature development tools. This established software base is a major reason teams have built their AI work around Nvidia hardware.

Code written and tuned for CUDA does not automatically move to another platform. Developers may need to rewrite kernels or replace software libraries, then check both the accuracy of their models and the speed at which they run. Training loops, the repeated steps through which models learn, also need to be adjusted for different compiler stacks. Compilers translate software into instructions the hardware can execute.

Chinese chip vendors offer their own development tools and connections to PyTorch and TensorFlow, widely used frameworks for building AI models. Compatibility and performance have improved, but gaps remain for some models, especially those that depend on custom CUDA extensions.

The domestic requirement will push more teams through this migration work. Progress in local hardware does not remove the need to test drivers, software tools and individual model operations. Matching Nvidia across performance, the range of supported software and driver maturity will take time.

What Determines an AI Cluster's Performance?

AI accelerators speed up the mathematics at the center of deep learning. Large training runs distribute that work across many chips, so the useful performance of a facility depends on more than the speed of any single processor.

Three factors matter most: raw computing throughput, memory bandwidth and fast connections between chips. Throughput describes how much calculation a processor can perform. Memory bandwidth determines how quickly it can receive the data needed for those calculations. Interconnects allow many processors to exchange information and operate together as a cluster.

Nvidia has strong positions in all three areas through advanced packaging, stacked memory and fast links that keep large clusters synchronized. Packaging concerns how components are assembled into a chip system, while stacked memory places memory layers together to support the flow of data needed by demanding AI workloads.

Domestic suppliers also face manufacturing constraints. Leading AI processors depend on advanced production processes and high bandwidth memory. Export controls on chipmaking equipment restrict access to some tools and materials needed to produce them.

Chinese factories can make capable devices, but production yields and energy efficiency often lag the newest overseas parts. Yield is the share of manufactured chips that work as intended. Lower efficiency can mean using more electricity and floor space to complete the same training work within the same time. Those requirements increase the cost of computing capacity.

Construction Schedules and Supply Chains Face Pressure

Many builds are expected to face delays in the near term. Procurement teams must renegotiate contracts, while system integrators need to qualify management software, drivers and orchestration layers for the replacement chips. Orchestration software assigns computing work and coordinates the machines carrying it out.

Different accelerators also have different electricity and cooling requirements. Operators will need to revisit their power and thermal budgets, checking whether planned electrical equipment and cooling systems can support the revised configuration. A hardware change can therefore affect both installation work and the operating plan for a completed facility.

Domestic accelerator lead times remain long, and production must expand to absorb demand previously served by foreign suppliers. A protected market creates a larger opportunity for Chinese vendors, but those vendors still need to deliver enough working equipment and support it at the scale required by major data center projects.

Gray market supplies that once helped fill gaps are being squeezed by tighter inspections and serial number tracking. System vendors will need to increase qualification work and customer support to keep projects moving as operators change equipment and purchasing arrangements.

Mixed Fleets Could Keep Services Running

Some operators may run a mixture of foreign and domestic hardware during the transition. Existing foreign accelerators could continue handling established inference services or research, while new training work moves to Chinese chips. Such arrangements would let operators assign different tasks to different parts of their computing fleet.

Mixed fleets also create technical complexity. Distributed training depends on rapid communication across many computing nodes. Differences in interconnects and compiler stacks can reduce efficiency when a single job spans equipment from multiple vendors.

Operators will therefore need to consider both chip availability and the way workloads are organized. Moving an individual service to a new platform involves different checks from running a large training job across an entire cluster. The migration combines hardware replacement, software validation and decisions about which machines should work together.

Domestic suppliers must improve the total cost of ownership, meaning the cost of buying and operating systems over their useful life. A performance gap can require more racks and more electricity for the same amount of work. That can increase operating expenses and slow the pace at which researchers expand model sizes and training runs compared with peers using leading GPUs overseas.

Growth Plans Shift Beyond China's Server Rooms

US chip suppliers will lean more heavily on demand in the United States, Europe, the Middle East and Southeast Asia. Some are investing in India and other markets to offset lost sales to projects backed by public money in China. Investors will assess whether those regions can absorb the capacity previously intended for Chinese data centers.

Chinese suppliers gain a protected home market, but the opportunity comes with pressure to expand production quickly and close software gaps. The policy moves demand toward domestic chips while leaving vendors responsible for making those chips practical and competitive across large deployments.

The directive also ties geopolitical decisions more closely to the construction of data infrastructure. A compliance decision about one accelerator can affect the schedule of a national AI program or a cloud service. Purchasing strategies, colocation plans and partnerships across borders will need to reflect hardware rules that reach into the equipment choices of strategic computing projects.

For affected operators, the immediate priorities are concrete: review government support, establish construction progress, audit equipment and orders, and prepare domestic replacements. The central change is that public support now carries a requirement governing the chips used to build new AI capacity.

Key Points

• China's new guidance requires new data center projects receiving state funding to use domestic AI chips.
• Projects less than 30 percent complete must remove installed foreign accelerators or cancel planned purchases; more advanced projects face individual reviews.
• Government support includes grants, tax incentives, discounted land, electricity price assistance and direct equity stakes.
• The restrictions cover Nvidia's H20 and target H200 and B200 processors that have reached buyers through gray market channels.
• AI data center projects have attracted more than $100 billion in public funding since 2021.
• Nvidia's Chinese accelerator market share fell from about 95 percent in 2022 to zero.
• Huawei, Cambricon, Enflame, MetaX, Moore Threads and Biren stand to gain demand.
• Replacing hardware requires software migration, system qualification and reviews of power, cooling and networking.
• Domestic supply constraints and differences in performance may increase costs and delay projects.
• Cybersecurity and economic planning regulators are reviewing compliance, with new capacity required to rely on domestic silicon.