China Drafts $295 Billion AI Infrastructure Plan, Mandating 80% Domestic Chips to Lock Out Nvidia

American export controls on advanced AI chips were designed to slow China’s AI development. The $295 billion plan taking shape inside China’s National Development and Reform Commission represents a systematic response: if you can’t have our chips, build your own infrastructure around different ones — and bake that preference into state policy at a scale large enough to be irreversible.

The NDRC, China’s top economic planning agency, is finalizing a blueprint to spend roughly 2 trillion yuan ($295 billion) over five years constructing a nationwide network of AI data centers. State carriers China Mobile and China Telecom will operate the bulk of the facilities and connect them into a unified computing grid by 2028. The plan has not been formally published and remains in drafting — specific targets may shift before final approval — but Bloomberg reported its outlines in June 2026, and the strategic logic is clear enough to analyze regardless of the final numbers.

The 80% Mandate

The most consequential technical specification in the plan is a requirement that at least 80% of AI chips and related hardware come from domestic suppliers. This is not a purchasing preference or a subsidy program — it is a national infrastructure blueprint mandating domestic-content thresholds into the physical foundation of China’s AI compute stack.

The domestically approved chip ecosystem has grown substantially. Nine categories of Chinese-developed AI accelerators — from Huawei’s Ascend line, Alibaba’s Hanguang series, Shanghai Biren Technology’s BR100, and Moore Threads — recently cleared a Chinese government security review for deployment in sensitive sectors. This review created a formal approved vendor list that the NDRC plan can now reference as its sourcing foundation.

Huawei’s Ascend 910C, the most capable chip on that list, has been closing the performance gap with Nvidia’s H100 generation — though independent benchmarks suggest it still trails the H100 on raw training throughput and is meaningfully behind Nvidia’s current H200 and B200 products. The gap in raw capability is real; whether it is large enough to matter for the kinds of AI workloads that will define the 2026-2028 period depends on what those workloads actually require.

Why This Structural Move Matters More Than the Dollar Amount

The headline figure — $295 billion, or $59 billion annually — is large but not startling in the context of global AI infrastructure spending. The major US hyperscalers (Microsoft, Google, Amazon, Meta) collectively committed over $700 billion in AI infrastructure investment for 2026 alone, and Saudi Arabia’s Project Transcendence announced a $100 billion AI data center commitment for the same period.

What matters more than the absolute size is the structural design. By building a nationwide compute grid with a domestic-content mandate baked into the blueprint, China is doing something that ad-hoc procurement preferences cannot accomplish: it is creating demand certainty for domestic chip makers at scale, over multi-year horizons, backed by state funding that does not depend on market economics.

This matters for Huawei’s roadmap specifically. Chip development is an extraordinarily capital-intensive process with 5-7 year lead times. Without guaranteed demand for Ascend chips in volumes large enough to justify the manufacturing investment, domestic development cannot achieve the learning curve effects needed to close the gap with TSMC-manufactured Nvidia products. The NDRC plan creates that demand in a way that individual procurement decisions cannot.

“The domestic-content requirement is embedded in a national infrastructure blueprint backed by state funding, making it harder to reverse than a procurement preference, even if U.S. export rules were loosened in the future,” noted one semiconductor industry analyst.

The G7 Context

The plan’s emergence coincides with ongoing G7 discussions about the US Fable 5 export control directive and related AI technology restrictions. China’s announcement of $295 billion in domestic AI infrastructure at the same time G7 nations are debating how to contain Chinese access to frontier AI models sends a clear strategic signal: the infrastructure gap that export controls are designed to preserve is being addressed through industrial policy, not circumvention.

A senior executive at a leading Chinese AI company was quoted stating that his company would match Fable 5-class model capability before Elon Musk’s Q1 2027 prediction for the same milestone. The $295 billion infrastructure plan is the hardware foundation for that ambition.

The Energy Equation

The compute buildout cannot be separated from the energy question. AI data centers are extraordinarily power-intensive — training runs for frontier models can consume hundreds of megawatt-hours, and inference at scale demands reliable, high-density power supply.

The NDRC plan includes provisions for renewable energy and nuclear power integration. If the power grid is included in the full investment scope, the total capital requirement could exceed 5 trillion yuan — transforming this from a data center project into one of the largest infrastructure investments in human history. China is the world’s largest installer of both solar panels and nuclear capacity, and both technologies are central to the energy strategy supporting this compute buildout.

What It Means for the Global Race

For US and allied AI companies, the $295 billion plan has two implications worth distinguishing.

The first is competitive: a unified national computing grid of this scale would give Chinese AI labs access to pooled compute resources in ways that are difficult to replicate through fragmented data center procurement. DeepSeek’s efficiency breakthroughs on the V3 and V4 models showed that Chinese labs are capable of extracting more performance per FLOP than the compute gap would suggest — a nationwide compute network makes that efficiency approach even more powerful.

The second is structural: by mandating domestic chips and locking in domestic infrastructure at this scale, China is making the technological decoupling between US and Chinese AI stacks largely permanent. Even if US export control policy were relaxed in a future administration, the procurement infrastructure, vendor relationships, and technical specialization built around domestic chips over the next five years would create enormous switching costs.

The export controls that were intended to create a window for US AI development to pull ahead may be simultaneously accelerating China’s development of a parallel compute stack robust enough to sustain frontier AI development independently. Whether that outcome serves US strategic interests is a question that policymakers in Washington are now actively debating — five years and $295 billion after the decision point has already passed.