China’s $295B AI Data Center Plan Squeezes Out Nvidia

14.06.2026 5 Min Read

China commits $295 billion to build a national AI data center network with an 80% domestic chip mandate, effectively barring Nvidia and positioning Huawei as the primary hardware provider.

On June 9, Bloomberg reported that Chinese government agencies have finalized a blueprint to spend approximately 2 trillion yuan ($295 billion) over five years on a nationwide network of interconnected AI data centers. The plan, coordinated by the National Development and Reform Commission (NDRC), is the largest single-government AI infrastructure commitment ever proposed. It comes with one condition that will reshape the global chip market: at least 80% of all hardware must come from domestic suppliers.

That requirement effectively bars Nvidia and AMD from competing for what could be the world’s largest equipment procurement contract. Instead, Huawei stands to become the central hardware provider for China’s sovereign AI cloud.

What the plan covers

The 2 trillion yuan figure covers computing infrastructure only. When electricity grid upgrades and power transmission facilities are included, the total investment climbs to roughly 5 trillion yuan ($737 billion), according to people familiar with the drafting process who spoke to Bloomberg. That sum exceeds twice the combined annual capital expenditure of all four major US hyperscalers — Meta, Microsoft, Amazon, and Alphabet — in 2025.

State-owned telecom operators China Mobile and China Telecom have been designated as the primary operators of the new facilities. The choice is deliberate: it keeps the infrastructure under direct government oversight rather than in the hands of commercial cloud providers like Alibaba Cloud or Tencent Cloud, whose own capital spending falls outside the 2-trillion-yuan envelope.

By 2028, the plan envisions disparate regional computing facilities — many already under construction — linked into a single coherent national network. Think of it as a sovereign cloud, disaggregated across the country but programmable as a unified resource for AI training and inference workloads.

Why Huawei wins, and Nvidia loses

The 80% domestic sourcing mandate is the single most consequential provision. Huawei’s Ascend series of AI accelerators — particularly the Ascend 910C — has been the primary beneficiary of China’s push to replace Nvidia hardware since US export controls tightened in October 2023. Huawei shipped an estimated 1.5 million Ascend chips in 2025, according to analysts at Bernstein, and the company has been aggressively courting cloud providers and research institutions.

The domestic sourcing clause transforms what has been a gradual, market-driven substitution into a procurement mandate. Every new data center built under this plan will be contractually required to fill at least 80% of its server racks with chips from Huawei, Cambricon, Biren Technology, or other domestic suppliers. Nvidia’s H20 — the downgraded chip it designed specifically to comply with US export rules — becomes irrelevant in this context. There is no reason to buy a restricted foreign chip when the government is mandating domestic alternatives anyway.

“This is not just import substitution — it is the construction of a parallel computing ecosystem,” said Paul Triolo, a partner at Albright Stonebridge Group who tracks China’s technology policy. “Beijing is betting that Huawei and its ecosystem can deliver enough performance to train frontier models, even if each individual chip trails Nvidia’s latest.”

The funding mechanism

The financial structure reveals how seriously Beijing treats this initiative. Sovereign debt instruments — including ultra-long-term special government bonds with maturities of 30 years or more — are expected to carry the bulk of the load. State-backed industry funds, commercial bank loans, and private investment will fill the remainder.

This is not unprecedented. China issued 1 trillion yuan in ultra-long special bonds in 2024, and an additional 1.3 trillion yuan in 2025, primarily for infrastructure and technology projects. The AI data center plan would require a comparable issuance on an annualized basis over five years, a pace that bond market analysts say is feasible given China’s current fiscal position but will add to sovereign debt pressures.

Supply chain implications

Upstream: The domestic chip sourcing mandate creates a massive captive market for Chinese semiconductor manufacturers. SMIC, which produces Huawei’s Ascend chips on its 7nm process node, will see demand surge. But SMIC’s capacity is already constrained — it is running its advanced lines at near-full utilization. The plan will accelerate capital expenditure at SMIC, Hua Hong Semiconductor, and other foundries, potentially diverting capacity from other customers.

Downstream: Chinese AI model developers — including Baidu, Alibaba’s DAMO Academy, ByteDance, and a wave of startups like Zhipu AI and Moonshot AI — will benefit from expanded computing resources. Training costs for large language models have been a bottleneck; additional domestic compute capacity could reduce the cost of training runs by 30-40% within two years, according to estimates from Morgan Stanley’s China technology team.

Bottlenecks: The plan’s success hinges on two chokepoints. First, SMIC’s ability to produce enough advanced chips at acceptable yields. Second, the availability of high-bandwidth memory (HBM) — a component where China remains heavily dependent on Samsung and SK Hynix. Beijing is investing in domestic HBM alternatives through CXMT and other memory chipmakers, but commercial-grade products are still 12-18 months away.

Market signal and trend assessment

Bull case: The plan accelerates China’s path to AI self-sufficiency. By 2028, a fully operational national computing network powered by domestic chips could support the training of models rivaling GPT-5 class systems. Huawei’s ecosystem matures, creating a viable alternative to the Nvidia-CUDA stack that dominates global AI infrastructure.

Bear case: Domestic chips still lag Nvidia’s Blackwell generation by 2-3 years in raw performance. The 80% mandate forces suboptimal hardware choices, increasing training costs and slowing iteration cycles. Chinese AI companies that need cutting-edge performance for frontier research will find workarounds — renting compute on overseas clouds, for instance — undermining the plan’s intent.

Base case: The plan delivers a meaningful expansion of China’s AI compute capacity, but the performance gap with the West persists. Chinese models improve steadily but rarely match the frontier. The real impact is structural: by 2028, China will have built a domestic AI hardware ecosystem large enough to sustain itself, reducing the leverage of US export controls as a policy tool.

What to watch: Huawei’s Ascend 910D chip, expected in late 2026, which will determine whether domestic silicon can close the performance gap with Nvidia’s B200. Also watch SMIC’s 5nm process development — if it achieves acceptable yields by mid-2027, the plan’s hardware targets become considerably more realistic.