China's Ministry of Industry and Information Technology published its 15th Five-Year Plan for the information and communications sector on September 7, 2026. This isn't a leaked draft or an offhand comment from an official at a conference. It's a formal five-year plan, the same category of document Beijing uses to lock in targets for steel output and rail mileage, and this one puts a hard number on AI computing capacity: 9,800 exaflops by 2030, up from a stated 2025 baseline of 1,590 eflops. Behind that target sits a commitment of 3.8 trillion yuan, roughly $532 billion, in cumulative information-infrastructure investment through the end of the decade.

What exactly did the plan commit to?

The headline is the compute target, but the plan gets specific about how China intends to hit it. It calls for the "orderly deployment" of large intelligent-computing clusters built around 10,000 GPU or accelerator cards, with some scaling past 100,000 cards. It also separates that training-scale infrastructure from dedicated inference-computing facilities, built for different application types rather than treated as an afterthought bolted onto training capacity. That distinction matters more than it looks on a policy page. It means Chinese planners are explicitly budgeting for the compute that serves AI models to users, not just the compute that trains them in the first place.

RelatedNvidia's $750B AI Deals Just Spooked Its Own Bondholders

How does that square with what China says it already has?

An exaflop is a quintillion floating-point operations per second, a unit that only means something in comparison. Going from 1,590 to 9,800 eflops is roughly a sixfold jump, though it's often rounded down to "fourfold-plus" in coverage of the plan, depending on which baseline gets used. Here's where the numbers get messy in an interesting way. At a State Council Information Office press conference on July 20, 2026, Xie Cun, director of the ministry's Information and Communications Management Bureau, said China's intelligent computing capacity had already reached 2,185 eflops at FP16 precision by the end of June 2026, up 177% year-over-year. That's higher than the plan's own 1,590 eflops "2025 baseline," and the two figures don't reconcile cleanly in the public materials. The likely explanation is a difference in measurement, precision standard, or scope of what counts as "intelligent computing." Both numbers come from the same ministry, months apart, and both are worth reporting as given rather than forced into one tidy line.

China's intelligent computing capacity: 2025 baseline vs. June 2026 actual vs. 2030 target Bar chart comparing three eflops figures: a 2025 baseline of 1,590 eflops, an actual reported 2,185 eflops as of June 2026 at FP16 precision, and a 2030 target of 9,800 eflops set by China's 15th Five-Year Plan for the information and communications sector. INTELLIGENT COMPUTE CAPACITY, EFLOPS China: reported baseline vs. actual vs. 2030 plan target 1,590 2025 baseline 2,185 Jun 2026 (FP16) 9,800 2030 target genztech.blog
Fig 1 China's own figures put compute capacity at 2,185 eflops by mid-2026, already past the plan's stated 2025 baseline, with more than a fourfold climb still needed to reach the 2030 target.

What does a 100,000-GPU cluster actually mean?

Clusters at that scale sit in the same conversation as the biggest AI infrastructure builds anywhere on the planet. A single cluster of 100,000 accelerator cards is hyperscaler-class hardware, the kind of buildout that in the US gets compared to Stargate-scale commitments from OpenAI, Microsoft, and their partners. Naming that scale explicitly in a national plan, rather than leaving it to individual companies to announce piecemeal, is itself a signal. Beijing is treating frontier-scale training clusters as national infrastructure, not just corporate capital expenditure, and pairing them with separate inference-serving facilities suggests the same training-versus-inference split now driving Western infrastructure debates, see Gimlet Labs and similar startups betting the real money is in serving models rather than training them, is showing up in Chinese planning too.

Can China get there without Nvidia and AMD?

This is the part of the plan that reads less like an infrastructure roadmap and more like a bet. US export controls have restricted the flow of advanced Nvidia and AMD chips into China for years now, and nothing in this plan suggests Beijing expects that to change before 2030. A national target calling for clusters of 10,000 to 100,000 cards, funded to the tune of $532 billion, is a statement about how much compute China believes it can secure through other means: domestic foundry capacity at SMIC, Huawei's Ascend accelerator line, Cambricon, Biren, and whatever else the domestic chip industry can bring to volume production over the plan period. Whether that supply chain can actually deliver enough capable chips at that scale is the question the plan doesn't answer, because no five-year plan can answer it in advance. It states the target and backs it with money. It doesn't guarantee the yield.

RelatedDeepSeek Pauses $1.5B Raise After Leaked Compute Remarks

  1. Jul 20, 2026State Council Information Office press conference: China's intelligent computing capacity reported at 2,185 eflops (FP16), up 177% year-over-year.
  2. Sep 7, 2026MIIT publishes the 15th Five-Year Plan for the information and communications sector, setting the 9,800-eflop, $532 billion target.
  3. 2030Target date for 9,800 eflops of intelligent computing capacity and the associated cluster and inference-infrastructure buildout.

What it means for the market

The plan draws a fairly clean line through the chip sector. Nvidia and AMD are structurally locked out of the bulk of this specific buildout, export controls make sure of that, so the direct beneficiaries of a $532 billion infrastructure push are domestic Chinese chipmakers and foundries: SMIC on the manufacturing side, Huawei's Ascend line, Cambricon, and Biren on the design side, plus whichever equipment and materials suppliers still have legal room to sell into China. The signal for investors watching this space is that Beijing's own central planning document now assumes, and funds, a compute buildout that runs mostly around Western hardware rather than through it. That's a data point about where the money is expected to flow, not a recommendation about where it will land.

What to watch: whether China's domestic chip yields and output can scale fast enough to fill 10,000-to-100,000-card clusters on schedule, whether Washington tightens or loosens export controls in response to a plan this explicit, and whether the training-versus-inference split written into this plan shows up in other countries' national compute planning the way it's already reshaping private infrastructure spending in the US.

Our take

A five-year plan is a statement of political will backed by a budget line, and it's worth treating it as exactly that, no more and no less. Compute buildouts are notoriously hard to forecast five years out. Chip yields, model architectures, and the entire economics of training versus inference can look different by 2028, let alone 2030. What makes this plan harder to wave off than a typical aspirational target is the money attached to it. $532 billion in committed infrastructure spending is not the kind of number a government puts on paper casually. Whether China lands on 9,800 eflops precisely is almost beside the point. The plan is evidence that Beijing has decided its answer to export controls is to out-invest them domestically, and that bet is now written into the country's central planning document for the next five years.

Primary sources

Original analysis by GenZTech Team. Source: State Council Information Office.