In March, China Telecom and Huawei reported an experiment linking three computing centres across Guangzhou and Shenzhen, spanning 409.61 kilometres. Training performance exceeded 97% of the centralised baseline in the reported test. It is an important demonstration of what high-capacity optical connections can make possible, although the workload, equipment costs and production scalability still need independent scrutiny.
The role in the system
Scale-up links, optical systems, wide-area networks and private networking.
Better links can improve useful system output; vendor claims still need production validation.
Companies in this layer
Huawei · China Mobile · VNET / DYXnet · ByteDance
Names identify research roles, not a claim that every company has a partnership with every other company.
What would strengthen the thesis?
Evidence of measured efficiency, reliability and interoperable workloads.
What would weaken it?
Technical progress that fails to improve usable output or economic returns; delays, costs or commercial friction that prevent the value test being met.
Supply-chain shipments versus useful system throughput
Kinwong’s preliminary October update attributed part of its growth to AI-server and high-speed communications board shipments. This is upstream production evidence, not verified Chinese-only deployments or measured end-to-end accelerator performance.
25–26 SEPTEMBER 2026 · EVIDENCE EXTENSION
Cross-region compute is changing the physical network — and the transport layer remains an active build problem.
CCTV reporting from Kaifeng, Henan shows backbone fibre routes being rebuilt on shorter paths because legacy routes laid for ordinary data transmission were not designed around today’s compute-latency requirements. This is direct physical evidence that a national compute fabric can require network-topology adaptation, not merely logical interconnection.
The 26 September review adds CAICT’s latest carrying-capacity work. The industry agenda includes expanding city-level 1ms compute coverage and piloting high-speed lossless transmission, task-based scheduling and compute-network agents. That reinforces interconnect as a first-class architecture layer between resource visibility and practical workload mobility.
Boundary: route upgrades and industry pilots do not establish nationwide low-latency coverage or make every remote workload viable. Latency, bandwidth, reliability, data-transfer cost and software portability remain workload-specific constraints.
Architecture notes
Interconnect
Scale-up, scale-out and scale-across solve different distance problems. Optics move closer to silicon as electrical paths become expensive; geographic distance ultimately requires scheduling rather than shared-memory metaphors.
Geographic compute fabric
At long distance the abstraction changes from one machine to one resource pool: compute becomes discoverable, measurable, connected and schedulable across operators and regions.