As the most decisive move in its strategy of technological self-sufficiency, the Chinese government has established a directive requiring the exclusive use of domestically manufactured artificial intelligence chips in new state-funded data centers. Confirmed by sources familiar with the matter, this applies to projects in planning stages as well as to facilities under construction and also affects those with any form of public involvement, from subsidies to equity participation or purchase commitments.
The operational detail sets the tone: since October, regulators have ordered projects with less than 30% progress to remove all already installed foreign AI hardware or cancel ongoing orders. For the more mature projects, the decision will be made on a “case-by-case” basis, but the underlying message is clear: critical AI infrastructure funded by public money will be built with domestic silicon.
This move closes the door to NVIDIA, AMD, and Intel—the major Western providers of AI acceleration—at public data centers, and includes “blocked” products for China, such as NVIDIA H20. It coincides with a particularly tense moment: after Washington confirmed that the latest Blackwell GPUs will not be exported to China, Beijing raises the stakes and accelerates its “domestic tech first” agenda.
What changes from today: no more foreign GPUs in publicly funded data centers
The directive is not limited to new tenders. According to available information:
- State projects with < 30% progress
Must remove already deployed foreign accelerators or cancel pending orders. The replacement must be domestic hardware (e.g., Huawei and other Chinese designers). - More advanced projects
Will undergo “case-by-case” evaluation. Practically, a default restriction is expected, with exceptions justified only when no viable domestic alternative exists in the short term. - Scope
Includes GPU/ASIC AI chips, accelerator cards, systems, and specific orders (including models tailored to China, such as NVIDIA H20).
The expulsion does not automatically bar private centers without public funding or purely commercial hyperscalers; however, the political signal could also favor domestic hardware in those segments, especially where sensitive data or services for the government are involved.
Why now: technological tensions and the “strategic silicon doctrine”
Beijing’s move must be read alongside the U.S. decision to ban the export of Blackwell GPUs (NVIDIA’s latest accelerators) to China. In this context, chips, memory, and interconnection are considered strategic assets, on par with energy or defense. The U.S. aims to maintain dominance in advanced semiconductors; China responds by accelerating substitution.
Impact for Western suppliers
- NVIDIA: loses another route into the Chinese public market. After dropping from about 90% to almost 0% market share in one quarter due to cross-restrictions, this directive consolidates that void. Models tailored for China (like H20) will not be available in public data centers; only private demand and the “grey market” remain as limited escape valves.
- AMD and Intel: are similarly excluded from state projects. Their options—like NVIDIA—are limited to private clients, unsubsidized academic collaborations, and entidades sin financiación pública.
Local beneficiaries
The counterbalance is clear: Chinese manufacturers such as Huawei (Ascend), Cambricon, MetaX, Moore Threads, or Enflame gain market share in a captive market. The directive guarantees demand, catalyzes investment, and pressures the ecosystem to mature hardware and software around local runtimes (replacements for CUDA), domestic compilers, libraries, and compatible toolchains.
Immediate costs: reduced performance and steep learning curves
China supports its industry but cannot ignore current disadvantages:
- Performance gap: local accelerators still don’t match the efficiency and performance of Western leaders in training frontier (large models, long context windows, token-intensive pipelines).
- Software stack: replacing CUDA involves rewriting parts of the AI stack (runtimes, libraries, optimizations). The challenge is both technical and organizational (talent, compatibility, reproducible performance).
- Manufacturing bottlenecks: sanctions restrict access to advanced nodes, machinery (lithography, metrology), materials, and state-of-the-art packaging. This limits the leap to cutting-edge technologies and raises production costs.
In summary: China can fill the gap with domestic volume, but at the expense of efficiency and TCO in the short/medium term. Training on the latest hardware will be more expensive and slower; inference is where local silicon may bridge the gap sooner.
Real-world impact: what will happen in data centers
- Project reshuffling
Early-stage government data centers will need to reschedule orders and redesign architectures (rack, power, cooling, networking). Those at 30% or less progress face sunk costs (installed hardware to remove) and renegotiations with suppliers. - Temporary halt on frontier deployments
Large-scale training with H100/H200 or Blackwell disappear from the public map. They might migrate to private services, grey chains, or overseas capacity, but not with state funding. - Boost for domestic purchase
Huawei Ascend and other local manufacturers will fill the pipeline. Integrators will need to adjust networks, frameworks, and libraries to domestic runtimes, with longer optimization plans. - Differentiation by levels
A stratification model is likely to emerge:- State (100% domestically produced)
- Mixed with public intervention (predominantly local)
- Purely private (greater freedom but political/sector pressure to “buy local”)
What about the “grey market”? Volume limits and compliance risks
The informal channels—intermediaries, triangulation, reselling—will persist, but the volume needed to train frontier models is hard to hide. Compliance has been reinforced in U.S. and allies, and large tech companies have incentives to avoid risking their licenses. In China, these routes may mitigate shortages for SMEs and medium labs; they do not replace public centers or national farms.
Who is most affected (and who can adapt better)
- NVIDIA: loses a key public market and sees limited the strategy of “capped editions” (H20). It will compensate using partner orders and AI factories in the West and Asia, but cedes China.
- AMD/Intel: face similar impacts in public data centers. They can seek to position themselves in private, edge, general-purpose PC/server, and x86/Arm chips for auxiliary services.
- Chinese providers: expand market share and knowledge through demand from public sectors. They must accelerate software stacks, optimize cost/efficiency, and multiply partnerships with local hyperscalers.
What to expect in 12–24 months
- Growth of domestic silicon in inference and public sector AI services
- Pilots and mixed services (private) with Western accelerators where there is clear ROI and legal coverage
- More regulations in China to deepen the use of domestic technology in critical infrastructure (AI, 5G, government cloud)
- Domestic “V2” accelerators with better efficiency and more mature stacks; replacements for CUDA with improved performance on popular models (translation, vision, speech)
- Divergence of standards: two less interoperable AI ecosystems
Key to the local ecosystem: from captive market share to excellence
To transform the directive from merely guaranteeing market share to becoming a quality driver, the Chinese ecosystem must:
- Invest in software: toolchains, frameworks, and libraries optimized (compilers, kernels, schedulers)
- Attract talent: engineers in systems, compilers, and machine learning focused on reproducible performance
- Enhance packaging: domestic capabilities in 2.5D/3D and interconnection to compete with HBM and NVLink in bandwidth and latency
- Increase transparency: measure and publish performance and efficiency on real workloads (RAG, fine-tuning, multi-tenant serving) to build market trust
Conclusion: self-sufficiency at any cost (for now)
Beijing’s decision toughens the tech decoupling, locks in China’s public sector reliance on domestic chips, accelerates local development, and sacrifices—for the moment—efficiency and cost compared to Western leaders. Simultaneously, it weakens NVIDIA, AMD, and Intel’s access to a high-volume and high-margin segment. In the short term, China will pay more for less performance in frontier training; in the medium term, if the domestic ecosystem matures software and packaging, it will close the gap in inference and services. In any case, the AI landscape—chips, memory, interconnection, and energy—is already a matter of industrial policy. And it will remain so.
Frequently Asked Questions (FAQ)
Which data centers are affected by the foreign AI chip ban?
The new state-funded data centers in China and any project with public intervention (subsidies, participation, commitments), even if under construction. Projects with < 30% progress must remove foreign hardware or cancel orders; the more advanced ones will be evaluated case by case.
Which manufacturers are excluded and which will benefit?
NVIDIA, AMD, and Intel are excluded from public projects; local firms such as Huawei (Ascend) and others (Cambricon, MetaX, Moore Threads, Enflame) will gain market share and knowledge.
Does the regulation affect private data centers without public funding?
The directive targets public-funded projects. In the private sector, companies are not obligated by this rule, but the political and sector pressure may tilt purchases toward domestic silicon, especially where sensitive data or public services are involved.
Will Western hardware still arrive via “grey market” channels?
Units might enter through informal channels but not in enough volume for frontier training of public projects. Regulations and compliance risks limit this route to occasional cases in the private sector.
Sources
- Sector guidelines and reports cited by specialized media about the mandatory use of domestic AI chips in state-funded data centers in China, including the withdrawal or cancellation of foreign orders in projects with < 30% completion.
- Previous coverage of the NVIDIA Blackwell export ban to China by the U.S., mentioning tailored products like NVIDIA H20.
- Public information about the Chinese AI accelerator ecosystem (Huawei Ascend, Cambricon, MetaX, Moore Threads, Enflame) and the limitations imposed by sanctions (equipment, advanced nodes, packaging).