While the West argues about which chatbot is cleverest, China’s Ministry of Industry and Information Technology has published a 2026-2030 plan to more than quadruple national AI computing capacity to 9,800 exaflops (a 1 followed by 18 zeros or a billion billion) by 2030. This serves as yet another reminder that the AI contest people can actually see, the model launches and the demos, sits on top of a far larger contest very few are paying attention to in the mainstream media: who controls the raw compute.
The same week, France’s Mistral raised a $3.5 billion war chest to stay in a race increasingly defined by capital and silicon rather than clever ideas. Compute is becoming the strategic commodity of the decade, closer to oil than to software, and the countries treating it that way are the ones building state-scale plans while everyone else looks on from the sidelines cheering ‘their team’ on. Nail-biting stuff that will likely define the remainder of this century.
What the plan actually commits to
The headline number is the quadrupling: China wants national computing power to reach 9,800 exaflops by 2030, more than four times its current capacity, with a heavy tilt toward the intelligent (AI-optimised) compute that trains and runs large models. The document is not a single data centre or a single company; it is a national industrial blueprint, the kind of five-year framework China uses to point capital, provinces and state-owned enterprises at the same target. Treat the exact figure as a stated ambition rather than a delivered fact, because plans and outcomes are different things. But the direction is unambiguous, and the scale is enormous.
The strategic logic is straightforward once you see compute as infrastructure rather than gadgetry. Export controls have restricted China’s access to the most advanced Western AI chips, so the plan leans on building domestic capacity at volume: more homegrown accelerators, more data centres, more power. If you cannot buy the best shovels, you make a great many of your own and out-dig the problem. It is a bet that in an AI race, aggregate national compute matters at least as much as any single frontier model.
Mistral, and why the timing rhymes
The Mistral news lands in the same key. Europe’s leading AI lab pulled in $3.5 billion, not because it lacks clever people, but because staying in this game now requires the capital to secure chips, energy and data-centre capacity at scale. When a well-regarded frontier lab has to raise billions mainly to buy its way to enough compute, that tells you where the binding constraint has moved. The ideas are necessary; the silicon and the electricity are what actually gate you.
This is the shift worth internalising. For a decade the AI story was about talent and algorithms, the scarce genius in the room. Increasingly it is about industrial inputs: fabrication capacity, power grids, cooling, and the geopolitics of who can access all three. That reframes the whole contest. It is less a science-fair rivalry between models and more an infrastructure build-out on the scale of railways or the electrical grid, with nation-states as the primary players and private labs increasingly dependent on state-level decisions about energy and chips.
The takeaway
Watch the compute, not just the demos. The model launches are the visible, shareable, argue-about-it-on-social layer, and they matter. But the deeper contest, the one that will shape which countries and companies can even afford to compete by 2030, is being fought over exaflops, fabs and power stations. China has now written its number down in a state plan. The honest question for everyone else is whether they are building at that scale, or just watching the leaderboard and cheering.
Related on Top Tool Stack: the Nvidia valuation debate, the company at the centre of the compute race.