China’s central government has issued a directive ordering all state-funded data centers to cease procurement of foreign AI chips and to begin phasing out existing Nvidia and AMD hardware, according to reports from Reuters and CNBC. The policy represents the most aggressive step yet in Beijing’s campaign to eliminate reliance on Western semiconductor technology in critical national infrastructure, and it goes further than previous guidance by setting concrete removal timelines rather than merely discouraging future purchases.
The directive applies to data centers that receive state funding or operate under state-owned enterprises, a category that encompasses a substantial portion of China’s AI computing infrastructure. Private companies with state affiliations — including units of ByteDance and Alibaba — have reportedly been told to halt testing of new Nvidia chips in any projects connected to government contracts or state-backed research initiatives.
The 50% Electricity Subsidy: Making Domestic Chips Commercially Viable
To ease the transition, the government is offering a significant financial incentive: data centers that achieve 100% reliance on domestically produced AI chips will receive a 50% discount on electricity costs. For large-scale AI training facilities, electricity is typically the single largest operational expense, often exceeding hardware amortization costs over a multi-year period. A 50% reduction in power costs is therefore not a marginal benefit. It is a structural competitive advantage that could offset the performance gap between domestic chips and Nvidia’s leading hardware.
This subsidy design reflects a sophisticated understanding of the economics of AI infrastructure. Rather than mandating compliance through penalties alone, Beijing is making the domestic chip ecosystem financially attractive enough that operators have a positive incentive to switch. The approach mirrors the subsidy structures used to build China’s solar and EV industries, both of which achieved global dominance through sustained state support during the early adoption phase.
Implications for Nvidia’s China Revenue and the Domestic Chip Ecosystem
Nvidia has already seen its China revenue decline sharply following successive rounds of US export controls that restricted sales of its most advanced chips. The new Chinese directive will further erode the market for whatever Nvidia products remain available in China, as state-affiliated buyers, historically a major customer segment, are now explicitly barred from procurement. The company’s H20 chip, specifically designed to comply with US export rules while remaining competitive in the Chinese market, may find its addressable market significantly reduced.
For domestic chip makers like Huawei, Biren Technology, and Cambricon, the directive is a windfall. It creates a guaranteed, captive market of large-scale buyers who must source domestically, providing the revenue and deployment scale needed to accelerate product iteration. As Alibaba launches its 10,000-Card AI Computing Cluster on proprietary hardware, the ecosystem of domestic AI infrastructure is already taking shape — the new directive will accelerate that trajectory across the entire state sector.
A Policy Shift That Redraws the Map of Global AI Infrastructure
The ban on foreign AI chips in state-funded data centers is a structural, long-term policy shift rather than a temporary measure. It signals that Beijing has concluded that the risk of dependency on foreign semiconductor technology outweighs the short-term performance benefits of using Nvidia hardware. The directive will reshape procurement decisions, investment flows, and talent strategies across China’s AI sector for years to come, and it marks a decisive step toward the bifurcation of global AI infrastructure into distinct Western and Chinese ecosystems. The scale of this shift, affecting not just purchasing decisions but the entire software and tooling ecosystem that Chinese AI developers use, will take years to fully materialize, but the direction is now unambiguous.
The Software Stack Problem: CUDA Dependency and the Cost of Migration
One dimension of the foreign chip ban that has received less attention than the hardware question is the software migration challenge it creates. China’s AI development community has spent years building tools, libraries, and workflows on top of Nvidia’s CUDA programming framework, which has no direct equivalent in the domestic chip ecosystem. Migrating existing codebases to Huawei’s CANN framework or other domestic alternatives requires significant engineering effort and introduces compatibility risks that can slow research and product development. The government’s 50% electricity subsidy helps with the financial cost of the hardware transition, but it does not address the engineering cost of the software migration. Chinese AI labs and data center operators will need to invest heavily in retraining their engineering teams and rebuilding their tooling infrastructure, a process that industry insiders estimate could take two to three years to complete at scale. The government has so far not publicly addressed this dimension of the transition, but the engineering cost is real and will need to be factored into any honest assessment of the policy’s total impact.