China targets 9,800 EFLOPS by 2030: your inference supply

China's Ministry of Industry and Information Technology published its five-year plan for the information and communications sector, and the headline is a compute target: 9,800 eflops of intelligent computing capacity by 2030. As of June 2026 the country was at 2,185 eflops. That is a more than fourfold increase in four years, funded by 3.8 trillion yuan (about $532 billion) of cumulative infrastructure investment. If any part of your stack runs on cheap inference from a Chinese-hosted model provider, this plan is the supply curve underneath your bill.

What actually happened

The South China Morning Post reported the plan, which covers 2026 through 2030. The figures worth keeping:

• 2,185 eflops of intelligent computing capacity as of June 2026, up 177% year over year
• 52 intelligent computing facilities already running more than 10,000 accelerator cards each
• The plan calls for "orderly deployment" of clusters at the 10,000-card and 100,000-card scale, plus inference-specific facilities
• 3.8 trillion yuan in cumulative information infrastructure investment across the period

Note what this is: a capacity target and a spending commitment, not delivered hardware. Five-year plan targets are direction, and China's accelerator supply is still shaped by export controls it does not set.

Why state-funded compute matters for your business

Cheap tokens have a policy dependency. The per-million-token price you like from a Chinese-hosted provider is partly a function of subsidized capacity and domestic competition. It is not a market price that a US small business can count on for a three-year contract. We have already watched accelerator rationing throttle model vendors with no warning to their customers.

Open weights survive a policy change; a hosted endpoint does not. This is the practical case for preferring providers whose models you could run yourself if the endpoint disappeared. You probably will never self-host. The point is that you could, which is why we keep saying open weights are a hedge, not a discount.

Capacity growth does not automatically mean your price drops. Inference capacity being built for domestic demand is not capacity aimed at your workload. Do not model next year's AI budget on a price cut somebody else's five-year plan implies.

Do the boring residency check now. Know which of your features call which endpoint, in which jurisdiction, with what data attached. Most small teams cannot answer that in under an hour. That is the actual risk — not geopolitics, but not knowing where your customer data goes when a support ticket gets summarized.

Source: South China Morning Post.