China to Quadruple Its AI Compute by 2030: 9,800 EFLOPS and 3.8 Trillion Yuan
China has put a number on its ambition: 9,800 EFLOPS of AI compute by 2030, more than 4.5 times what it had in June. The MIIT five-year plan turns the model race into an infrastructure race.
What China Announced: the MIIT Five-Year Plan
China's Ministry of Industry and Information Technology (MIIT) published its 2026-2030 five-year plan for the information and communications sector in the week of September 7. It is the document that decides where the country's digital infrastructure gets built over the next five years, and its central goal is explicit: to take China's intelligent computing capacity to 9,800 EFLOPS by 2030. This is not a lab target, it is a deployment target with money behind it.
From 2,185 to 9,800 EFLOPS, More Than 4.5x
The starting point is what makes the number big. At the end of June 2026 China had 2,185 EFLOPS of intelligent compute, up 177% year over year. Going from there to 9,800 EFLOPS means multiplying installed capacity by more than 4.5 in roughly four and a half years. One EFLOPS is one trillion floating-point operations per second: the unit now used to measure how much training and inference a country can absorb.
3.8 Trillion Yuan in Information Infrastructure
The plan lays out 3.8 trillion yuan, about 566 billion dollars, in cumulative investment in information infrastructure through 2030. Data centers, power, cooling, fiber and edge compute. It is the least visible part of AI and the part that decides whether models can actually be trained and served at scale.
Read also
The Network: 5G, 5G-A and 95% Penetration
The same document sets connectivity goals for 2030: 50 5G base stations, including 5G-A, per 10,000 inhabitants, and a 5G user penetration rate of 95%. Without that network, distributed compute and edge inference never reach factories or cities.
100,000-Accelerator Clusters on Domestic Silicon
The most discussed part of the plan is architecture: compute clusters of 100,000 accelerator cards, alongside 10,000-card clusters, with explicit priority for domestic technology. In other words, AI supercomputers built with as little dependence on foreign suppliers as possible.
JD Cloud and Moore Threads GPUs
On September 9, JD Cloud said it plans to build a 100,000-GPU cluster using chips from China's Moore Threads, aimed at large-model training, inference and embodied intelligence, with capacity available to enterprises. It is the clearest example of the plan turning into actual purchases.
The Sugon 8000 Precedent
This does not start from zero. In July 2026 Sugon announced the completion of the Sugon 8000 (Dengfeng), presented as the first domestic production AI supercluster with 100,000 GPUs. The MIIT plan turns that experiment into industrial policy.
The Global Race: Japan, Europe and the United States
China is not playing alone. Japan plans to more than quadruple its AI compute by 2033 with roughly 60 billion dollars of investment. In Europe, Google announced 13 billion euros of AI infrastructure in Finland on September 9. And the United States keeps restrictions on advanced Nvidia chips in place, pushing Beijing toward domestic silicon. The debate is no longer which model is best, but where it runs and with what power.
What It Means for Developers and Companies
For anyone building software, the signal is twofold. First, more Asian compute supply can pressure inference prices and open alternatives to Western providers. Second, hardware fragmentation makes portable software more valuable: containers, vendor-neutral runtimes and exportable models matter more than ever. If your product depends on a single cloud, plans like this are a reminder to diversify.
FAQ
What Is an EFLOPS
It is a unit of computing power: one trillion single-precision floating-point operations per second. It is used to measure AI compute capacity at national scale.
How Much AI Compute Does China Have Today
2,185 EFLOPS at the end of June 2026, according to MIIT itself, after 177% growth in a year.
What Is a 100,000-Accelerator Cluster
A data center that interconnects 100,000 compute cards (GPUs or AI accelerators) to train large models and serve inference at scale.
Why Does China Prioritize Domestic Chips
Because of US export restrictions on advanced accelerators, which make self-sufficiency a condition for hitting its targets.
Conclusion
The message of China's plan is that the next competitive edge in AI is built with concrete, power and silicon, not algorithms alone. If you follow how each announcement reshapes the industry, keep reading the blog: the compute race has only just started. For more on the models themselves, see our breakdowns of DeepSeek V4.1-Flash and Mistral's record funding round.
