China set a target of 9,800 EFLOPS of computing capacity by 2030
The five-year plan from the Ministry of Industry and Information Technology is backed by 3.8 trillion yuan. It is a response to export controls that assumes they will not be lifted.
China's Ministry of Industry and Information Technology has set a target of 9,800 EFLOPS of intelligent computing capacity by 2030, backed by 3.8 trillion yuan — roughly $530 billion — according to the South China Morning Post.
The figure is a planning target in a five-year programme, not a procurement order, and Chinese industrial targets have historically been met unevenly. But the direction is unambiguous, and the assumption underneath it is the important part: the plan is built for a world in which US export controls stay where they are.
What the number means in practice
EFLOPS totals are difficult to compare across jurisdictions because the precision being measured is rarely specified consistently, and aggregate national capacity says little about whether it is concentrated usefully. A hundred thousand accelerators split across provincial data centres serving local government workloads is not a frontier training cluster.
The more concrete signal is the deployment underneath. DeepSeek is planning a cluster of more than 160,000 Huawei AI accelerators at a one-gigawatt site in Inner Mongolia, according to Bloomberg — among the largest known deployments of domestic Chinese silicon, and a direct test of whether Huawei's parts can sustain frontier-scale training.
That is the question the entire programme turns on. Chinese labs have shipped competitive models — Alibaba refreshed Qwen3.8-Max at 2.4 trillion parameters with a one-million-token context window this month, and Z.ai released GLM-5.3-Flash as a 320-billion-parameter open-weight multimodal model at a tenth of its predecessor's cost. Whether they can keep doing so on domestic accelerators at scale is unproven.
The comparison that matters
South Korea has committed $919 billion to an AI infrastructure programme targeting 8.4GW by 2029 and 18.4GW by 2035. The US hyperscalers are building past a gigawatt per campus, with Alphabet alone carrying $811 billion in infrastructure purchase commitments and OpenAI holding a partnership to deploy at least 10GW of Nvidia systems.
China's 3.8 trillion yuan over five years is comparable in magnitude to any of these. The difference is not capital. It is what the capital can buy: every other programme is spending into a market where the best accelerators are purchasable, and China's is not.
Why the workarounds do not change the plan
Restricted hardware still moves. Taiwan indicted nine people last month over 74 smuggled B300 servers, and the New York Times reported that Inspur shipped about $3 billion in Blackwell systems to Southeast Asia through its US subsidiary Aivres.
Neither is a foundation for a five-year industrial plan. Smuggled supply is unreliable, unsupported, unserviceable and legally precarious, and a subsidiary route can be closed with a single addition to the Entity List. A national programme cannot be built on a channel that a foreign regulator can shut in an afternoon.
So the plan assumes domestic silicon, which is why the Inner Mongolia cluster is the thing to watch rather than the headline number. If 160,000 Huawei accelerators can train a frontier model, the export control regime has bought time and nothing more. If they cannot, it has bought a durable advantage — and the 2030 target is a statement of intent about which of those China expects to be true.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters