On 8 September, China's Ministry of Industry and Information Technology presented a 2026-2030 plan aiming to bring national intelligent computing capacity to around 9,800 exaflops, up from 2,185 at the end of June and around 2,450 at the end of July. The plan provides for around 3.8 trillion yuan, or nearly 532 billion dollars, in cumulative investment in information infrastructure.
We have been tracking computing as a strategic resource for months. Here is the most explicit document published on the subject by any state.
The figure that makes the target credible
A five-year goal does not commit to much. What makes this one serious is the trajectory already covered: capacity is said to have grown by 177% in a year, reaching 2,185 exaflops at the end of June.
At that pace, the goal of quadrupling in four years is not unrealistic. It is even a conservative projection, which is unusual in this kind of document.
The announced amount is on the same scale as the cumulative US private commitments we document, except that this is a coordinated public plan rather than a sum of corporate decisions.
What this confirms
This plan validates a hypothesis we put forward two weeks ago, which remained a hypothesis.
We wrote, regarding GLM-5.3-Flash running on Chinese chips, that if confirmed, the question of catching up would change in nature: it would no longer be about models but about circumventing the hardware constraint.
A state plan aiming to quadruple national capacity implies accelerator production on a corresponding scale. It is a direct response to export controls, and it is now quantified and dated.
An important methodological caveat. The term refers to capacity optimised for AI, measured in units that depend on the chosen computing precision. Cross-country comparisons are therefore tricky: depending on whether you count in high or reduced precision, the same hardware can show very different figures. This plan indicates an ambition and an order of magnitude; it does not allow direct comparison with US or European capacity.
What this changes in the landscape
Three consequences emerge, and they overlap with what we observe.
The open strategy makes sense. We have documented the massive release of Chinese models with open weights, from Qwen to GLM under an MIT licence. A country building its own computing capacity does not need to sell access to its models: it can give them away and capture value elsewhere, in influence and ecosystem.
Export controls change their target. If they were meant to slow things down, they may have accelerated autonomy. This is a classic and documented effect of technology restrictions, and it will take several years to know which of the two logics prevails.
Third countries gain an alternative. We saw this with an Arab model built on Chinese weights. National computing capacity makes that offering more credible over the long term.
What to take away
This plan is interesting less for its figures than for what it says about the competitive arena.
Three weeks ago we wrote that AI had become an industry, with the physical constraints that come with it: electricity, land, components. A five-year infrastructure plan is exactly the kind of document produced for a heavy industry, not for a software sector.
And this poses a question to Europe that it can no longer avoid. Faced with a public plan of 532 billion dollars and US private commitments on the same scale, a 3-billion-euro raise and a supercomputer at 388 million do not belong to the same category. That does not mean giving up: it means choosing a strategy that does not consist of running the same race.