In October 2025 the General Office of China’s Ministry of Industry and Information Technology issued a notice launching a city-area “millisecond computing” initiative. The goal is clear: by 2027, form a city-area compute network with full coverage and efficient connectivity — millisecond interconnection between compute centres for infrastructure, and millisecond reachability for application terminals.
From “enough” to “fast enough”
Earlier compute policy mainly addressed whether there is enough compute: building centres, promoting east-to-west computing and increasing total capacity. “Millisecond computing” shifts attention to whether compute is fast enough — that is, whether users can obtain compute services with sufficiently low latency.
There is a clear technical background. Inference-side applications, especially real-time interactive ones such as agents, live translation and industrial control, are highly latency-sensitive. However large total capacity is, excessive scheduling latency cannot support such scenarios.
What a latency target means for policy
Writing latency into a policy objective means three things.
First, compute layout evolves from concentrated to layered. Training can be concentrated in the west, while inference needs to sit close to users, forming a centre-edge structure. See our note on building an integrated national computing system.
Second, compute services evolve from resource rental to service levels. Measurable latency will push compute services towards tiered products and service level commitments. See our note on the national computing network and data express routes.
Third, compute scheduling becomes a distinct technical and compliance issue. Cross-region scheduling involves data flow, identifier management and security monitoring, and is directly related to data compliance. See our note on network data security risk assessment.
What it means for AI companies
First, product design must consider latency constraints. Real-time applications need to weigh latency in compute architecture decisions, not just unit cost per unit of compute.
Second, compute procurement agreements need finer service level clauses, with latency, availability and incident response written in and backed by liability. See our note on AI commercial contract clauses.
Third, edge deployment brings new compliance questions. With inference nodes more widely distributed, data flow and retention across nodes requires a redesigned compliance approach. See our data compliance service.
Turning compute policy into measurable indicators is a sign of AI infrastructure maturing. For application-layer companies, compute choice has become an architecture question rather than merely a cost question.
This article is based on the MIIT notice on the city-area “millisecond computing” initiative and on public information. It is for general reference only and does not constitute legal advice.