China wants to make idle compute borrowable — like electricity

A research lab in Hangzhou finishes a training run and leaves a rack of GPUs dark for the weekend. Three hundred kilometers away, a team in a western hub is desperate for exactly that capacity. Today, those two machines might as well be on different planets. A 2025 plan wants to change that.

The idea is simple to say and hard to do: make compute flow. Not just build more of it, but let unused capacity be found, booked, and borrowed across owners, chip architectures, and regions — the way electricity is drawn from a grid rather than hoarded in private generators.

The plan and its two clocks

The document is the Action Plan for Computing Interconnection — 《算力互联互通行动计划》 — issued by the Ministry of Industry and Information Technology (工业和信息化部) and published on 30 May 2025. It sits on top of an earlier, broader directive: the 2023 opinion Deepening the "East Data, West Computing" Project and Accelerating the National Integrated Computing Network (发改数据〔2023〕1779号), which set the geographic and green-power frame. The 2025 plan is the connective tissue — the standards and platforms that make the physical build-out actually usable as one network.

The plan sets a two-step timeline:

• By 2026: establish a fairly complete system of computing-interconnection standards, identifiers, and rules. On the facility side, promote new high-performance transport protocols and raise inter-node network connectivity. On the resource side, build national, regional, and industry interconnection platforms, unify public-compute identifiers, and connect the public resources of the country's leading compute providers. On the business side, enable compute–storage–network service interoperability and cross-owner, cross-architecture, cross-region scheduling. On applications, run computing-internet trial-network pilots.
• By 2028: essentially achieve standardized interconnection of national public compute, and gradually form a computing internet with intelligent感知 (perception), real-time discovery, and on-demand access.

The phrase "like water and electricity" appears in official explainers: the end state is compute you draw when you need it, from wherever it sits.

How "interconnection" actually works

Three moves define the mechanism.

First, a unified identifier. Every participating compute resource gets a standard "compute identifier (算力标识)" — a tag that says what it is, where it is, and what it can do. Without a common label, a scheduler cannot match a job to a machine it has never met. The plan runs a "畅联" (free-connect) campaign to push unified identifiers into use.

Second, layered platforms. It calls for a national computing-internet service platform (国家算力互联网服务平台), plus pilots of M regional and N industry interconnection platforms, and a hub-interconnection center to strengthen coordination between national hub nodes. The national layer aggregates identifiers; the regional and industry layers solve local and sector-specific matching pain.

Third, business interoperability. The plan wants compute, storage, and network services to interoperate, so a task can be scheduled across owners and architectures — including China's mixed domestic and foreign chip lines. That is the hardest part, because heterogeneous hardware rarely speaks the same scheduling language out of the box.

The five campaigns

To execute, the plan names five campaigns (五大行动):

1. Compute-interconnection foundation building — publish a technology catalog, build the standards system, and set up a compute-market evaluation framework.
2. Facility-interconnection speed-up — upgrade internet backbones and exchange centers, expand carrier bandwidth, and cut transmission latency.
3. Interconnection-platform system building — stand up the national service platform, run the identifier "畅联" campaign, and pilot regional and industry platforms plus the hub-interconnection center.
4. Business-interconnection innovation — trial new scheduling and access services, and list benchmark firms by service type.
5. Compute-interconnection navigation — lead on standards and promote the "China solution" internationally.

Underneath sit six work areas and 16 tasks, covering core-technology breakthroughs (high-speed interconnect buses, high-performance transport protocols, identifier gateways, compute–network–cloud scheduling), standards, interconnection rules (including compute metering, pricing, and trading), facility links, resource sharing, business interoperability, application scenarios, and security safeguards.

Why idle compute is the target

The problem the plan names is utilization, not just supply. China has poured capacity into western hubs, yet public compute from different providers often cannot be discovered or booked by outside users. Expensive machines sit idle while other teams scramble for access. Interconnection is the lever that turns fragmented, underused capacity into a liquid market.

The plan dovetails with the data-labeling and computing-voucher policies covered elsewhere on this site: a schedulable national network is what makes those vouchers redeemable against capacity outside your own city. Subsidy and interconnection move together.

The hard parts nobody can skip

Three obstacles are real. Heterogeneity — mixing architectures means translation layers and shared protocols that do not yet exist at scale. Trust and settlement — cross-owner scheduling needs metering, pricing, and settlement rules the plan only begins to sketch. Security — a network that lets any authorized job run anywhere is also a network with a larger attack surface, so the plan pairs interconnection with compute-network and data-security safeguards.

None of these is solved by the document alone. The 2026 milestone is deliberately modest — standards and identifiers in place — leaving the harder "actually runs as one grid" work for the 2028 stretch goal.

Honest limitations

This article is built on the MIIT-issued Action Plan for Computing Interconnection (published 30 May 2025) and official explainers from provincial cyberspace and industry bodies, cross-checked against China News Service, Workers' Daily, and China National Radio reporting. Caveats:

• The plan is attributed to the Ministry of Industry and Information Technology in the most direct authoritative reporting (China News Service, 30 May 2025); some secondary outlets referred loosely to "five departments," but the issuing authority cited in primary coverage is MIIT. The plan itself builds on the 2023 five-department "East Data, West Computing" opinion (发改数据〔2023〕1779号).
• The 2026 and 2028 targets are policy goals, not achieved outcomes. The 2026 bar is specifically about standards, identifiers, and rules being "fairly complete" — a process milestone, not a working grid.
• The "M regional and N industry platforms" are framed as pilots; the plan does not, in public summaries, fix exact numbers for M and N.
• This piece focuses on the 2025 interconnection plan and deliberately does not re-litigate the 2022 hub construction or the 2023 integrated-network opinion, which are covered separately on this site.
• No RMB investment figure is stated in the public summaries of this plan, so no currency conversion is applied.

What readers can do now

1. If you buy GPU or cloud capacity in China, watch the national computing-internet service platform and the regional/industry pilot platforms — a functioning identifier-and-scheduling layer would let you book western-hub capacity you cannot reach today.
2. If you plan AI workloads, design jobs to tolerate heterogeneous hardware and remote scheduling now; the plan's business-interoperability goal points toward cross-architecture task placement becoming normal.
3. If you track infrastructure policy, use the 2026 standards/identifier milestone as the public scoreboard — that is when the plan's promises become checkable, well before the 2028 "one grid" goal.