China just budgeted 532 billion dollars to stop needing Nvidia

Nine thousand eight hundred exaflops. That is the intelligent computing capacity China's Ministry of Industry and Information Technology wants the country running by 2030, according to the 2026 to 2030 industry plan it published on 8 September.

The starting point is 2,185 exaflops at FP16, measured at the end of June 2026. That figure was itself up 177 percent year on year. Quadrupling it again inside four and a half years is the headline ambition, and the money attached is 3.8 trillion yuan, roughly 532 billion dollars, in cumulative information infrastructure investment.

What the China AI compute capacity 2030 plan actually contains

The country already operates 52 intelligent computing facilities carrying more than 10,000 accelerator cards each. The plan asks for orderly deployment of clusters at that 10,000 card scale and, separately, at 100,000 cards and above.

Those are not abstract targets. A 100,000 card cluster is a specific engineering problem involving power delivery, interconnect and cooling at a scale only a handful of organisations anywhere have solved. Writing it into a national industry plan is a statement about which problems the state intends to fund through to completion.

For scale on the private side, Anthropic alone has committed to 517 billion dollars of compute deals. China's figure is a national programme covering five years and a broader set of infrastructure, so the two are not directly comparable, but they sit in the same order of magnitude.

The investment figure covers information infrastructure broadly rather than accelerators alone. Networking, data centre construction and 6G work sit inside the same envelope, so the compute portion is a fraction of the 532 billion dollars and not the whole of it. The constraint that eventually binds is electricity, and the power draw of AI data centres is already the harder number to source.

The line that matters is about chips, not exaflops

Further down the document sits a request for greater efforts to adapt infrastructure to domestically produced computing chips.

That sentence does more work than the capacity target. It separates a plan that assumes access to Nvidia hardware from one that assumes the opposite. Adapting infrastructure to domestic silicon means software stacks, interconnect standards, cluster scheduling and training pipelines all built against parts that are not H100s and will not behave like them.

It also explains a detail that reads as a footnote on its own. The PyTorch Foundation added Alibaba Cloud and Cambricon as Platinum members on 8 September, with Ant Group joining at Gold. Cambricon makes domestic AI accelerators. A governance seat inside the framework most of the world trains on is what you want when your hardware needs first class support rather than a compatibility layer.

The same thread runs through hardware launch order. Arm's Mali G2 Ultra NX shipped in China first, which tells you something about where the volume is expected to be.

Why the estimate needs revising

Almost every public projection of Chinese AI capability has assumed a hardware ceiling set in Washington. Export controls limit the top end parts, and capability tracks available compute.

A five year programme designed around domestic silicon does not break that logic, but it changes the input. The question stops being how much restricted hardware gets through. It becomes how fast domestic accelerators close the gap per watt and per dollar, and how much of the software ecosystem moves to meet them.

The plan answers neither. Plans state intent, and Chinese industry plans have a mixed record on hitting their numbers. The 2,185 exaflops already built and the 52 facilities already running are the part that is not a projection.

So: the number to watch is not 9,800 exaflops in 2030. It is what fraction of the next 10,000 card cluster ships with domestic accelerators inside it, and whether anyone publishes the utilisation figures afterwards.