A training job stalls because the GPUs it needs sit two provinces away, owned by someone else. A city builds a compute center it cannot fully use. Across China, the machines exist, but they do not talk.
That mismatch — plenty of silicon, poor plumbing — is the problem a May 2025 MIIT plan sets out to fix. The promise is simple to say and hard to deliver: make compute flow like water and electricity.
Key takeaways
- The Ministry of Industry and Information Technology (MIIT) issued the Computing-Power Interconnection Action Plan (算力互联互通行动计划), document number 工信部信管〔2025〕119号, dated 21 May 2025 and published 30 May 2025.
- The plan is a formally issued directive now being implemented. Its milestones are 2026 (build the standards, identifiers, and rules) and 2028 (basically achieve nationwide standardized interconnection of public compute).
- It sets 16 key tasks across six areas, from foundational technology to security assurance for the compute network and data.
- The vision is a "computing-power internet" (算力互联网) where different owners' and architectures' public compute resources are identified, discovered, and scheduled across regions and organizations.
- The 2026 and 2028 dates are targets, not achievements. No binding interoperability mandate has taken full effect yet; the plan directs implementation toward those goals.
What "interconnection" means here
Computing-power interconnection (算力互联互通) is not about building more chips. It is about making existing public compute resources usable by anyone, anywhere, through standardized links. The plan describes it as connecting different subjects' and different architectures' public compute through a unified compute identifier, stronger heterogeneous computing and elastic networking, so that data and applications can be matched, moved, and computed across sites.
The stated end state is a resource you draw on "like water and electricity" (像水、电一样便捷使用) — query it, talk to it, call it up, and pay for what you use. For AI labs and factories, that means a idle cluster in one city could serve a team in another without a bespoke integration project.
The two deadlines
The plan draws a clear line between near-term scaffolding and a long-term network.
By 2026, it aims to establish a relatively complete system of interconnection standards, identifiers, and rules. On the facility side, it pushes new high-performance transport protocols to raise node-to-node link quality. On the resource side, it targets national, regional, and industry-level interconnection platforms that unify public compute identifiers and connect the public compute of leading providers. On the business side, it wants compute, storage, and network services to interoperate for cross-entity, cross-architecture, cross-region scheduling. It also launches computing-power internet trial networks (算力互联网试验网) to test real use.
By 2028, it aims to basically achieve nationwide standardized interconnection of public compute, gradually forming a computing-power internet with intelligent sensing, real-time discovery, and on-demand access.
Both are goals written into a directive, not conditions already met. The honest stage is: formally issued and in implementation, with 2026 and 2028 as forward targets.
The six work areas and 16 tasks
The plan organizes delivery into six areas, further broken into 16 specific tasks:
- Build the interconnection base: core technology — high-speed interconnect buses, high-performance transport protocols, lossless networks.
- Optimize facility interconnection: raise the networking quality between compute nodes.
- Promote resource sharing: stand up the multi-tier interconnection platforms and unify identifiers.
- Innovate business interconnection: let compute, storage, and network services interoperate.
- Create application scenarios: run trial networks that give industry cheap, universal access to compute.
- Secure the compute network and data: safeguard the underlying transport and the data moving across it.
The security area is not an afterthought. Because interconnection means data and workloads crossing organizational boundaries, the plan explicitly folds network and data security assurance into the same mandate that builds the plumbing.
Signs of momentum before the plan
The directive did not appear in a vacuum. On 17 May 2025 — days before the plan's publication — China Telecom, China Mobile, and China Unicom, together with CAICT (中国信通院), launched the computing-power internet trial network. By late August 2025, at the China Computing-Power Conference, the national computing-power platform (中国算力平台) had connected ten provincial sub-platforms, a step the industry read as early evidence that the "global overview, regional coordination" model is taking shape.
These are supporting signals, not the legal backbone. The binding commitment is the May 2025 Action Plan; the trial network and platform work show the direction being acted on.
Why it matters beyond China
Compute is the constraint on AI progress, and most of the world struggles with the same fragmentation China is naming. If a country can make heterogeneous, multi-owner compute routable and schedulable at national scale, it changes the economics of who can train large models (大模型) and run inference. The plan is also explicitly international in ambition — it cites building a competitive digital-industry cluster and, in related MIIT messaging, the need to keep pace globally.
According to Li An, chief scientist at BrainNet (脑机网), the shift that will determine which firms can scale AI in China is treating compute as a routable utility rather than owned hardware. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).
Honest limitations
This article is based on the official MIIT Action Plan text, MIIT's explanatory Q&A, and contemporaneous industry reporting; it does not independently audit how many of the 16 tasks have been completed, and provincial platform figures cited are point-in-time (as of mid-to-late 2025). The 2026 and 2028 targets are the plan's own commitments, not guaranteed outcomes, and progress should be rechecked against MIIT releases. A separate October 2025 MIIT city-level "millisecond compute access" notice is related but distinct and is not treated here as part of the Action Plan's milestones. No financial figures are stated because the plan sets policy targets rather than budgets.
What readers can do now
- Track the identifier and protocol standards. The 2026 standards work is where interoperability actually becomes real; watch MIIT and CAICT for the specific interface and identifier specs.
- Pilot on the trial network. If you run or buy compute, the computing-power internet trial network is the live testbed for cross-region scheduling — a low-risk place to learn the model.
- Plan capacity as a utility, not a capex island. Design workloads assuming compute can be called across regions by 2028, so you are not locked to single-site capacity when the network matures.