---
title: "China's scattered compute, one network: the interconnection deadline"
date: 2026-11-15
category: Policy & Governance
site: NeuroAI
canonical: https://neuroai.site/a/na-policy-compute-interconnection-plan
language: en
---

# China's scattered compute, one network: the interconnection deadline

> The MIIT's Compute Interconnection Action Plan (21 May 2025) sets 2026 as the year China's public compute must interconnect, full interconnection by 2028.

Consider a Chinese province where a robotics startup waited six weeks for GPU time while, a thousand kilometers away, a provincial computing center ran its clusters at half load. The machines could not talk to each other. China has built enormous compute capacity, but much of it sits in silos owned by different firms, provinces, and architectures that do not interoperate.

A plan published in May 2025 is meant to end that waste. The Compute Interconnection Action Plan (算力互联互通行动计划), issued by the Ministry of Industry and Information Technology (MIIT) as document 工信部信管〔2025〕119号 and dated 21 May 2025, tells operators to make public compute resources interconnect (互联) under common standards, identifiers, and rules.

## A plan, not a penalty law

The first thing to get right is the stage. This is a published action plan (行动计划), circulated with the instruction to "implement it in light of actual conditions (请结合实际认真抓好落实)." It is in the rollout phase: pilots, platform construction, and standard-setting are underway. It is not a binding regulation with fines; it is a coordinating blueprint with deadlines.

That matters for readers. An action plan signals direction and unlocks budgets and pilot zones, but compliance is expressed through the standards and platforms it commissions, not through a courtroom. Treat it as "in implementation," not "mandatory law."

## The 2026 and 2028 targets

The plan follows a stated sequence — pilot before scale, interconnect before network (先试点后推广、先互联再成网) — so the 2026 milestone is explicitly a foundation, not a finish line.

The plan is built around two clocks.

- By 2026: establish a relatively complete system of standards, identifiers, and rules for compute interconnection (算力互联互通标准、标识和规则体系). Build national, regional, and industry-level interconnection platforms, unify public compute identifiers, and achieve interconnection of the public compute resources of the country's leading compute enterprises.

- By 2028: basically achieve nationwide standardized interconnection of public compute, and gradually form a "compute internet (算力互联网)" with intelligent sensing, real-time discovery, and on-demand access.

The 2026 line is the near-term test. If the standards and the first platforms land on schedule, the country's top compute providers should be able to route work to each other's idle capacity through a common addressing scheme.

## What "interconnection" actually requires

The plan names six aspects and sixteen key tasks (六方面16项重点任务). The hard technical work breaks into three layers.

- Facility interconnection: adopt new high-performance transport protocols and raise the network link quality between compute nodes.

- Resource interconnection: build the national, regional, and industry platforms and assign every public compute resource a unified identifier, so a job can find spare capacity by address rather than by phone call.

- Business interconnection: let compute, storage, and network services be scheduled across owners, architectures, and regions — the part that actually moves a task from the idle cluster to the waiting startup.

It rests on earlier groundwork: the "East Data, West Computing" (东数西算) implementation opinion (发改数据〔2023〕1779号) and the High-Quality Development Action Plan for Computing Infrastructure (工信部联通信〔2023〕180号). Interconnection is the next layer on top of those build-outs.

## Who pays, who complies

The plan is addressed to the communications-management departments (通信管理部门), industry authorities, and the three state telecom carriers (China Mobile, China Telecom, China Unicom) plus China Broadcasting Network and China Tower, along with "relevant units." In practice, the large cloud and intelligent-computing operators are the ones expected to expose public capacity through the common identifiers and platforms.

According to Li An, chief scientist at BrainNet (脑机网), standardized compute interconnection could let smaller brain-computer labs borrow idle training capacity across provinces instead of buying their own clusters — a direct benefit for research groups whose workloads are bursty rather than constant.

## Why it matters for AI

Large model (大模型) training is the obvious beneficiary: unified addressing means a model developer is no longer hostage to the one operator that happens to have free cards. But the deeper point is strategic. Compute (算力) is now treated as infrastructure on par with power and transport — something the state should be able to see, route, and allocate, not just something firms buy.

That framing carries a question the plan does not answer: how voluntary interconnection stays voluntary as it becomes essential. The 2026 milestone will show whether leading providers genuinely open capacity or merely register it.

## Honest limitations

This article describes the plan as published on 21 May 2025 and in implementation, with 2026 and 2028 targets; the precise content of the supporting standards and the membership of the interconnection platforms may evolve and is not fully public. My account draws on the MIIT-issued text as reproduced by government portals and on corroborating industry reporting. I have not verified which specific enterprises have already interconnected, nor the current utilization rates of provincial centers. The "leading compute enterprises" the plan references are not individually named in the document, so I avoid listing particular firms as confirmed participants.

## What readers can do now

- If you run AI workloads in China, watch the 2026 standards and identifier scheme — they will determine whether you can burst into another provider's idle public compute instead of over-provisioning.

- Track the national and regional interconnection platforms as they come online; early access may matter more than raw card count for bursty training jobs.

- Read the plan as a signal of compute (算力) being treated as state infrastructure — plan your capacity strategy around interconnection, not just ownership.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-compute-interconnection-plan
Free to quote with attribution and a link to the original.
