9 SEP 2026 — China's Ministry of Industry and Information Technology has published its fifteenth five-year plan for the information and communications sector, targeting 9,800 exaflops of intelligent computing by 2030 against 2,185 at the end of June, with ¥3.8 trillion of cumulative infrastructure investment. Read as a growth rate rather than a multiple, the target is a sharp deceleration from what the country has just done.
The figures, and the one most coverage drops
The plan was published on 7 September, dated 12 August, and covers 2026 to 2030. Its 2025 baseline is 1,590 exaflops. Intelligent computing capacity reached 2,185 exaflops by the end of June 2026, up 177 per cent year on year, and the 2030 target is 9,800.
The ¥3.8 trillion figure — about 532 billion dollars — is cumulative information infrastructure investment across the period, not a compute budget.
The detail that keeps falling out of the coverage is the unit. The June figure is measured at FP16, a reduced numerical precision. That is the right convention for AI workloads and it is not the convention supercomputer rankings use, so an exaflop here and an exaflop in a Top500 listing are different quantities by roughly an order of magnitude.
Quadrupling is the slow option
Going from 2,185 to 9,800 is a multiple of about 4.5, and the period from mid-2026 to the end of 2030 is about four and a half years. Compounded, that is roughly 40 per cent a year.
The year to June 2026 delivered 177 per cent. The plan therefore describes growth at somewhere under a quarter of the rate just achieved, which is not how a fourfold increase reads in a headline.
Take the plan's own 2025 baseline of 1,590 instead and the implied rate is about 44 per cent a year over five years. The conclusion does not move: on either starting point, the target is a deceleration.
Three readings, and they are not equivalent
The first is that the plan is conservative by design, as five-year targets often are, set low enough to be met and reported. That is the most common explanation for a state planning number and it is probably part of the answer.
The second is that 177 per cent was a one-off. A year of aggressive procurement ahead of tightening export controls, followed by a slower steady state, would produce exactly this shape — a spike in the base year and a modest forward rate.
The third is the uncomfortable one: the planners expect supply to constrain them. If the accelerators available domestically cannot sustain the recent rate, then 40 per cent a year is not caution, it is arithmetic.
Domestic silicon is the variable
The plan is described as built on domestic chips as United States export controls tighten, which makes the target a forecast about Chinese fabrication as much as about spending.
That is where the three readings separate. A conservative target and a supply-limited one look identical in a document and diverge entirely in what they imply about capability by 2030, and no published figure distinguishes them.
What would is a breakdown of expected domestic against imported accelerator contribution, which the plan does not give and which no ministry would be expected to publish.
Orderly deployment is a brake
The plan calls for orderly deployment of clusters with more than 10,000 accelerator cards, and the adjective is doing real work. It signals central coordination over where capacity is built, after a cycle in which provincial construction produced data centres in the wrong places at poor utilisation.
The cluster threshold points the same way. Specifying scale says fragmented capacity does not count toward the goal, because training runs need concentration and a hundred scattered installations do not aggregate into one usable cluster.
Both are corrections to a previous overbuild rather than accelerants, which fits a plan whose headline growth rate is lower than its recent history.
The number the plan is not denominated in
Compute at this scale is an electricity commitment with a computing application, and the target is expressed in flops. Power planning sits in other instruments, which is normal, and it means the compute plan alone cannot be checked for feasibility.
We have watched that constraint bind elsewhere this month — an interconnection queue frozen at 474 gigawatts in Texas, and a gigawatt campus announced in Hyderabad with no grid connection date. China's grid position is stronger than either, and the question is the same one.
The figure to track is not the 2030 target but the next quarterly capacity print. If it comes in near the recent rate, the plan was cautious drafting. If it comes in near 40 per cent, the ceiling is silicon.
