China's Ministry of Industry and Information Technology published its 15th Five-Year Plan for the information and communications industry on 7 September 2026, and one figure travelled further than the rest: a target of 9,800 EFLOPS of intelligent computing capacity by 2030. The plan pairs that goal with a call for 3.8 trillion yuan, about 532 billion US dollars, in cumulative information-infrastructure investment across the 2026 to 2030 period. It is one of the most explicit national compute targets any government has set, and reading it well means understanding both what the number counts and what it quietly leaves out.
What "intelligent computing" actually measures
The headline says 9,800 exaflops. An exaflop is a billion billion floating-point operations per second, and for two decades the word anchored to the TOP500 supercomputer rankings, which measure double-precision (64-bit) math used in physics and climate simulation. AI does not run on that arithmetic. Training and inference lean on much lower precision, 16-bit, 8-bit, increasingly even 4-bit numbers, because neural networks tolerate the imprecision and run far faster for it.
"Intelligent computing capacity" is measured in those lower-precision operations. That is not a trick; it is the honest unit for AI work. But it does mean the figure is not comparable to a traditional supercomputer's FP64 rating, and headline EFLOPS numbers from different sources can differ by an order of magnitude depending on the precision assumed. When you read that China aims for 9,800 EFLOPS, the useful mental translation is "roughly four and a half times its current AI-optimized throughput," not "9,800 times a Frontier-class supercomputer."
The useful translation is roughly four and a half times current AI throughput, not thousands of times a traditional supercomputer.
On reading intelligent-compute figures
A roughly four-and-a-half-fold climb
China's intelligent-computing capacity as of June 2026 against the 2030 target, in EFLOPS. These are lower-precision, AI-optimized operations and are not comparable to traditional FP64 supercomputer ratings.
The target implies a fast but decelerating build relative to the 177 percent year-on-year growth reported to mid-2026.
That framing matters because the growth rate is the real story. Capacity reached 2,185 EFLOPS by the end of June 2026, up 177 percent from a year earlier. A target of 9,800 by 2030 implies sustaining a fast but decelerating build, not an impossible one. The plan is ambitious in absolute terms and plausible in trajectory terms, which is exactly why it is worth taking seriously rather than dismissing as a slogan.
The constraint the plan does not name
The obvious question a compute target raises is: on whose chips? The most capable AI accelerators are designed by United States firms and manufactured in Taiwan, and successive rounds of export controls have restricted China's access to the highest-end parts. A national target of 9,800 EFLOPS is therefore also, implicitly, a target for domestic silicon. Meeting it largely on home-designed accelerators would require Chinese chipmakers to scale production and yields dramatically over four years.
This is the tension running under the whole document. A plan to quadruple AI compute is only as real as the supply of accelerators to fill the buildings, the power to run them, and the memory to feed them. China has reported building 52 intelligent-computing clusters each with more than 10,000 accelerator cards, and the plan calls for the orderly deployment of clusters at the 10,000-card and 100,000-card scale. The physical footprint is already large. Whether the domestic supply chain can deliver frontier-class throughput at that volume, under export restrictions, is the unresolved variable that determines whether 9,800 is a floor or a ceiling.
Central target versus distributed capex
The plan is also a useful contrast in how two systems are financing the same race. China is setting a top-down national number and directing investment toward it. The United States is reaching comparable scale through private capital: hyperscaler capital expenditure and large private data-center commitments that, added together, run into the hundreds of billions of dollars a year without any single government target coordinating them.
The plan in four concrete commitments
The document pairs a headline compute target with an investment envelope and a build program for large accelerator clusters.
A central target buys coordination; whether domestic silicon can fill the clusters under export controls is the open variable.
Neither approach is obviously superior, and the honest position is that they optimize for different things. A central target buys coordination, the ability to align power grids, land, and chip production toward one goal, at the cost of the risk that the target is set wrong and capital is steered into stranded capacity. Distributed private capex buys adaptability and market discipline at the cost of duplication and concentration in a few firms. The 532 billion dollar figure also covers all information infrastructure, not AI alone, so it is not a like-for-like comparison with private AI capex; it is a signal of state priority rather than a precise AI-only budget.
Why the target exists at all
Behind the engineering is a straightforward strategic logic. Compute has become an input to economic and military capability the way electricity once was, and a country that depends on foreign suppliers for that input is exposed in exactly the way an energy-importing nation is exposed. The plan is an attempt to make AI compute a domestic utility rather than an imported good. That is the same instinct driving Europe's sovereign-AI push and the United States' own reshoring of chip fabrication, expressed through central planning rather than subsidy or private markets.
For everyone downstream of these decisions, the buildout has a quieter implication. Compute is consolidating into a handful of national and corporate blocs, each with its own chips, its own models, and its own rules about who can use them. The lesson that applies at the national level applies at the level of a single organization too: dependence on any one stack is a risk, and the ability to move work across providers and models is leverage. Platforms like Metir that stay model-agnostic express that principle at the software layer, and China's plan is the same argument about dependence playing out at the scale of a state.
The measured read is that the target is credible in trajectory and uncertain in supply. The growth rate that gets China to 9,800 EFLOPS is a deceleration of what it has already achieved, which makes the goal realistic on paper. What the plan cannot guarantee is the silicon, the power, and the memory to fill the clusters under export controls, and that gap, not the headline number, is where the plan will actually be decided.
Sources:
- China targets fourfold boost in AI computing capacity by 2030 in major tech push | South China Morning Post
- MIIT Plan Targets 9,800 Eflops of Intelligent Compute by 2030 | Unite.AI
- China releases 5-year development plan for information, communications sector | Xinhua
- Connecting nation's computing networks | China Daily
Image credits
Hero image: Beijing's central business district, by user Fzhenan1, via Wikimedia Commons, licensed under CC BY-SA 4.0. In-body photograph: a large server installation by NASA, public domain, via Wikimedia Commons. The server photograph is a generic illustration and does not depict a specific Chinese facility.