In May 2025 MIIT issued its Action Plan for Compute Interconnection. The operative word is not building but interconnecting: the plan addresses the problem that plenty of compute has been built, yet it is not always usable or smoothly usable.
From scale expansion to interconnection and scheduling
In recent years the emphasis in compute construction has been on totals — how many intelligent computing centres have been built, how many accelerator cards deployed. But compute is distributed across different operators, architectures and regions, and these do not connect, so local idleness and local scarcity exist side by side.
The plan’s aim is to link these islands into a network: promoting new high-performance transport protocols, raising the level of interconnection between nodes, and enabling cross-operator, cross-architecture and cross-region scheduling of compute supply and demand. See our note on the national integrated compute system and large models.
Public compute identifiers as the institutional pivot
The plan calls for a unified pool of public compute identifiers. Identifiers matter because they make compute a resource that can be searched and traded — with a common identity description, scheduling and settlement finally share a language.
This approach connects to earlier work on a compute internet service platform. See our note on the launch of national compute internet service platform routes.
Complementary to millisecond computing
Interconnection addresses whether things can be connected; latency governance addresses whether they run fast. The metropolitan millisecond computing campaign advanced in parallel fills in the performance metrics of the compute network from another direction. See our note on MIIT’s metropolitan millisecond computing campaign.
Where companies can start
First, watch the access rules for public compute identifiers, and assess whether the company’s own compute resources need to join the identifier system.
Second, move compute procurement from nearest to best price, since interconnection makes cross-region price comparison possible.
Third, assess the compliance cost of data moving with compute. Compute scheduling is often accompanied by data flows, and outbound transfer and entrusted processing rules need checking at the same time. See our notes on the action plan for building high-quality industry datasets and on State Council AI development and deployment.
In the next stage of compute policy, competition will not be about who has built the most but about who can organise compute efficiently.
This article is based on MIIT’s Action Plan for Compute Interconnection and public information and is for reference only; it does not constitute legal advice.