Beijing's draft plan funds a state run network of AI data centers and ties the buildout to a procurement threshold that, if enforced, removes Nvidia and AMD from China's state funded AI infrastructure market.
China's drafted 2-trillion-yuan (~$295B) plan to build a national network of AI data centers does not ban Nvidia by name. It sets a procurement parameter: 80% domestic content. That number is the policy.
The plan, reported by multiple outlets covering the draft document, commits roughly $295 billion of state-directed spending to a five-year buildout of a unified AI data-center network. If the 80% threshold is enforced as written, the rule does not just tilt the market. It removes Nvidia and AMD from China's central AI infrastructure by procurement math, not by export control.
This is the architectural difference from the US Stargate project. Stargate is private capital building private capacity. The Chinese version is state capital building a state network, with state carriers such as China Mobile and China Telecom acting as the network fabric and 2028 as the interconnection target. The scale reference is useful; the operating logic is not the same. The 80% domestic-share rule is written into the procurement design of a sovereign infrastructure plan, not the operating budget of a corporate consortium.
The mandate sits inside the "Eastern Data, Western Computing" framework, publicly outlined by China's National Data Bureau as four directions for an integrated national computing-power network. AI infrastructure is the headline bet, not chip self-sufficiency for its own sake. The 15th Five-Year Plan, submitted to the National People's Congress on March 5, 2026, names AI more than 50 times while barely foregrounding semiconductors. The policy bet is on compute capacity inside a connected grid, not on chips as a standalone industry.
That framing matters for what the 80% rule actually does. A chip export ban denies supply at the seller side and is enforced by one government against another. A domestic-content procurement mandate denies demand at the buyer side and is enforced by a state against its own operators. The lever is reversed. Any foreign accelerator, from Nvidia and AMD to Intel or a future entrant, that lands in the "AI chip" category becomes structurally excluded from the central state grid, even if every export-control license is intact. The number, not the company, is what the doctrine is aimed at.
The named beneficiaries are the domestic accelerator designers the rule is built to scale. Per coverage of the draft, the supply side is Huawei, Biren, Moore Threads, and Alibaba T-head. The procurement lever rewards whichever of those vendors can deliver volume, performance, and software stack at grid scale. It penalizes any roadmap that depends on imported accelerators for the central fabric.
The supply-side risk is the binding constraint, and it is the falsifier for the whole doctrine. The 80% threshold assumes the domestic chip base can absorb 80% of a national AI grid at competitive performance by 2028. That is a steep curve. China's H100-class supply is still scaling, advanced packaging capacity is constrained, and the gap between domestic accelerators and Nvidia's leading edge in raw training throughput is real, even if narrowed over the last 18 months. If domestic volume or per-chip performance falls short, the rule either softens in finalization, gets interpreted flexibly at the regional implementation stage, or forces Chinese AI labs to route around it, the way ByteDance already does.
That workaround dynamic is the second tell. ByteDance is routing 36,000 Nvidia B200 Blackwell chips through Malaysia via a roughly $2.5 billion deal with Southeast Asian cloud firms, exploiting a gap between US export-control scope (which is vendor and SKU based) and Chinese procurement scope (which is project and category based). The 80% rule closes the central grid to that routing. It does not close the private, regional, or extraterritorial routes that a major commercial AI lab can still build. The mandate is precise about which lanes it controls.
The conditional has to survive into the read. The plan is reported as drafted, not finalized. Enforcement mechanism, exact ratio interpretation, and timeline slippage are not yet fixed in the available record. Multiple outlets have corroborated the headline numbers, while the primary government text for the 80% mandate is not directly available in the reporting. If the ratio softens in finalization, the doctrinal story weakens. If domestic supply cannot bend to 2028, the rule becomes a procurement preference with a strong default, not a structural exclusion.
The watch item is supply, not geopolitics. The question for the rest of 2026 is whether Huawei, Biren, Moore Threads, and Alibaba T-head can collectively deliver 80% of a national AI grid at competitive performance on the 2028 schedule. If they can, the 80% rule becomes a permanent feature of how state buying power is wielded against any foreign chip category, and the next procurement doctrine will be aimed at the next layer of the stack. If they cannot, the doctrine reveals itself as a target the state set for its own chip industry, and the 80% number is a forcing function rather than a finished wall.