--- title: "China's computing-power blueprint sets a 300 EFLOPS bar — and a green one" date: 2026-10-09 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-policy-compute-infra-2025-action-plan language: en --- # China's computing-power blueprint sets a 300 EFLOPS bar — and a green one > The 2023 six-department Action Plan for High-Quality Development of Computing Power Infrastructure puts hard numbers on AI-era compute, storage and network targets through 2025, with a low-carbon clause built in. The data center down the road is quietly competing for the same power and land as a new AI lab. China's top industry regulators decided that competition needed a rulebook. So they wrote one — and put hard numbers on the table. ## What the plan actually sets out In October 2023, six departments — the Ministry of Industry and Information Technology (工业和信息化部), the Cyberspace Administration of China (中央网信办), the Ministry of Education (教育部), the National Health Commission (国家卫生健康委), the People's Bank of China (中国人民银行) and the State-owned Assets Supervision and Administration Commission (国务院国资委) — jointly issued the *Action Plan for High-Quality Development of Computing Power Infrastructure (算力基础设施高质量发展行动计划)*. It runs to 2025 and breaks the target into four buckets: 计算力 (computing power), 运载力 (network transport capacity), 存储力 (storage capacity) and 应用赋能 (application empowerment). The document is not a vision statement; it is a delivery sheet with 25 key tasks across six work streams. ## The numbers that matter By 2025, the plan calls for: - total computing power (算力) above **300 EFLOPS**, where EFLOPS means 10^18 floating-point operations per second; - intelligent computing (智能算力) — the GPU-heavy kind that trains large models (大模型) — reaching **35%** of the total; - storage volume above **1,800 EB**, with advanced storage at 30% or more; - optical transport network (OTN) coverage of **80%** at key application sites; - **40%** adoption of IPv6 and SRv6 innovative technologies on backbone and metropolitan networks; - **100%** disaster-recovery coverage for core and important data in key industries; - at least **30 application benchmarks** in each priority sector — industry, finance, healthcare, transport, energy and education. It also wants eastern and western regions to develop computing in a balanced way, which is the policy expression of the "east data, west computing" (东数西算) layout already taking shape across national hub clusters. ## Why storage and transport share the page Most people read "computing power" as chips. The plan treats it as a system: compute, network and storage have to move together. The 1,800 EB storage target and the OTN coverage rule exist because a model training job is only as fast as the slowest of the three. The 1.5x latency ceiling between national hub clusters is the transport counterpart to the 300 EFLOPS compute target — a number that means little without the pipes to feed it. ## The green clause is not decoration A full section is devoted to low-carbon computing (绿色低碳算力). The plan tells operators to lift the carbon efficiency of data centers, pull in green energy through source-grid-load-storage (源网荷储) techniques, and raise the on-site share of renewable electricity. For a country where data-center power demand is climbing fast, that clause is a constraint new builds must design around, not a slogan. The plan also names six special upgrade campaigns: computing supply, compute-network fusion, storage-compute coordination, compute application, green low-carbon, and security assurance. ## Who this actually moves The targets are binding on regulators and state-owned operators, and they shape procurement for anyone buying capacity at scale. A 35% intelligent-computing share means general-purpose clouds are being pushed toward AI-suitable hardware. An 80% OTN coverage rule means operators must wire priority sites to the national backbone. And the renewable-electricity push means power sourcing is now a compliance question, not just a cost question. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the plan's real signal is that compute is treated as national infrastructure on par with energy and transport, which changes how private AI labs should plan multi-year capacity. ## Honest limitations All figures above are policy targets stated in the official government release, not measured outcomes. The plan sets direction; provincial execution, grid constraints and chip-supply realities will determine how close 2025 numbers get to paper. Figures are government-disclosed; independent third-party audits of actual EFLOPS and renewable share are not yet published in a comparable format. This article is a policy explainer, not investment advice. ## What readers can do now - **Map your workload to the targets.** If you buy or lease compute in China, check whether your provider can show progress toward the 35% intelligent-computing share and the 1.5x inter-hub latency ceiling — these affect real training and inference latency. - **Ask for green evidence.** Request renewable-electricity share and carbon-efficiency numbers before signing a multi-year data-center contract; the plan makes this a stated policy direction, so providers should have an answer. - **Watch the published indices.** The plan mandates a national computing platform and periodic indices; track those rather than vendor marketing when sizing future capacity. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-compute-infra-2025-action-plan Free to quote with attribution and a link to the original.