Infrastructure implementation · 17 October 2026
Coordination is only real when a workload has power to run
The latest AI Plus discussion gives computing–electricity coordination a prominent place in national policy. That matters because moving a workload between regions requires more than a scheduling interface: the receiving facility must have available electrical capacity, a functioning network, suitable accelerators and a commercial arrangement that makes the transfer worthwhile. The 16 October State Council readout signals policy priority, not completed integration.
At VNET's Ulanqab campus, September reporting identified the unfinished Dongying 500kV project as an electricity-supply dependency. Its eventual commissioning is a concrete observation point. Even then, available power would establish only one link in the chain from a planned campus to billable, utilised computing capacity. Evidence of routine cross-region scheduling, cost savings and customer payments remains more persuasive than a national capacity total.
Evidence integration · 17 October 2026
Horinger offers a test of integrated deployment—not yet national dispatch
China Telecom’s Horinger campus is now associated with a proposed 10,000-card Huawei Ascend 950DT cluster and a newly topped-out 3D intelligent-computing building. The operator–equipment relationship is supported by the named project, and the location places the build within Inner Mongolia’s western computing geography. The reported 30,000-card provision is a design capability rather than installed capacity.
Regional reporting says electrical, cooling and associated systems have reached system-integration testing, with core construction targeted for late October. Acceptance, chip installation, measured throughput, customer contracts and cross-regional workload scheduling have not been established by that report. A functioning local cluster is an important prerequisite for a national computing network; it is not proof that capacity can already be dispatched commercially across provinces.
Control-plane development · 16 October 2026
Coordination requires permissions as well as scheduling
A national computing network cannot be judged only by whether workloads can move between clusters. It must also control which agents may access which data and services, provide reliable audit trails, and allocate responsibility when autonomous tasks behave unexpectedly. Research into agents running on Tencent Cloud and Tencent's subsequent acknowledgement of unintended browsing behaviour make that requirement concrete. The incident does not demonstrate a breach of China's national compute network, but it raises the standard of evidence for secure, interoperable orchestration. Read the underlying research.
Evidence from several stages of the network is beginning to accumulate. In a March demonstration, China Telecom and Huawei reported training across three computing centres in Guangzhou and Shenzhen at more than 97% of a centralised baseline. Wuhan, meanwhile, has reported RMB62.4 million in cumulative transactions through its computing marketplace. The first is a technical demonstration; the second is evidence of a market beginning to trade computing services. Neither yet shows that large volumes of paid training routinely move between provinces at attractive economics.
From visible capacity to purchasable compute
Shanghai’s Haina marketplace is a regional attempt to make computing services easier to buy. National cross-cluster software initiatives pursue interoperability, while 9 October central policy supports integration and warns against wasteful investment. Verified cross-region job volumes and cash returns remain open questions.
National design, market-selected locations, and the missing scheduling proof
Evidence on Inner Mongolia power economics and high-speed optical corridors suggests China's national strategy is best described as state-shaped infrastructure and geography with commercially selected nodes, not centrally allocated workloads. Regional network latency and power advantages can expand the set of viable remote workloads; they do not establish a unified national scheduler or measured cross-region goodput.
Reports of a proposed US multi-provider compute-grid consortium also complicate the simple China-versus-US dichotomy: distributed compute scheduling can emerge through both public coordination and private-provider federation. The relevant comparison is governance, interoperability and measured useful work, not the mere presence of a marketplace.
Proof required: live interregional dispatch logs, paid cross-provider transactions, repeatable workload performance, settlement and cost per useful token.
A western park has cheap electricity and substantial idle compute. Why can’t eastern AI demand simply be sent there?
Because available compute is not automatically usable compute. The July evidence describes exactly the gap a national coordination layer is trying to close.
Reported constraint evidence · Economic Observer · 19 Jul 2026: one Inner Mongolia park had annual utilisation below 50%; another heterogeneous domestic-compute pool averaged below 20%. The report attributes part of the latter problem to failure to uniformly manage different domestic architectures at the scheduling layer. It also reports 4.2–5.8 ms one-way Inner Mongolia-to-eastern-city latency and a manufacturing buyer saying more than 60% of its demand could not move west. These are attributed reporting and interview evidence, not a national utilisation census.
Read the source ↗12 SEPTEMBER 2026 · OPERATING COORDINATION EVIDENCE
So what is China doing about the fragmentation? Xiong’an New Area — a state-backed development zone in Hebei province — has put a cross-region computing coordination centre into operation.
Official Xiong’an and Hebei sources say the centre connects 44 compute resource pools across 19 provinces, covers all eight national computing hubs and aggregates more than 27,000P of compute. It combines a compute-dispatch platform with a model-service platform focused on token dispatch.
What this changes: the national-network thesis now has stronger operating-platform evidence. What it does not prove: that the connected capacity is interchangeable, that workloads are routinely moving across provinces at scale, or that dispatch produces attractive utilisation and cash economics.
30 SEPTEMBER 2026 · FINANCING LAYER
The national computing-power network is now explicitly inside a policy-bank support channel.
The PBOC has added 算力网 to the infrastructure categories eligible for PSL support. This strengthens the policy-and-finance side of national coordination: the network is increasingly being treated as strategic infrastructure. It does not show which projects will receive funds, how much will be deployed, or whether cross-provincial compute becomes more usable as a result.
Read the evidence and boundary →25–26 SEPTEMBER 2026 · VISIBILITY, NETWORK AND WORKLOAD-MOBILITY UPDATE
The national system is becoming measurable — and the steps between “visible”, “reachable” and “movable” compute are becoming clearer.
The National Data Administration says that at end-July China had 2.45 million PFLOPS (FP16) of intelligent compute, with 1.45 million PFLOPS — about 60% inside the national monitoring scope. By early September the trial platform covered 1,462 compute facilities. This materially strengthens the evidence for national resource visibility.
But CCTV reporting on 24 September says compute monitoring and cross-provincial invocation remain incomplete and many resources are still “islands”. In Kaifeng, Henan, backbone fibre routes are being rebuilt on shorter paths because older routes were not designed for today’s compute-latency requirements. Visibility therefore does not equal reachability, and reachability does not equal workload portability.
The 26 September review adds another transport-layer signal. At a CAICT forum, the 2026 compute-network carrying-capacity work emphasised expanding metro 1ms-compute coverage and continuing R&D/pilots in task-based scheduling and compute-network agents. That is useful evidence of direction and active implementation work — but also evidence that these capabilities are not yet universal.
Shanghai adds a stronger execution signal: a three-day electricity-price-aware inference test moved work among locations including Shanghai, Anhui and Xinjiang, while a wider 16-data-centre event coordinated flexible load, backup power and cross-region compute transfer. A Zhejiang test separately moved part of a Jiaxing workload to Qinghai. These examples demonstrate bounded workload mobility, not a nationally fungible compute pool.
National Data Administration ↗ CCTV via Digital China ↗ Xinhua / CAICT network-transport report ↗ State Grid Shanghai ↗ China Energy News ↗
23 SEPTEMBER 2026 · ARCHITECTURE UPDATE
National coordination now sits inside a nested optimisation model
The latest research links local system optimisation to the geographic layer: package and SuperPoD resources feed cluster/campus control, which can connect to regional and national scheduling, and—where workloads permit—to compute–electricity coordination.
What this adds: the national layer is not presented as a single giant cluster. It is a higher-level coordination domain whose value depends on interoperability, workload eligibility, network performance, energy conditions and economics.
See the full architecture →22 SEPTEMBER 2026 · ARCHITECTURE REFINEMENT
How a coordinated national system could bridge supply constraints
The architecture now brings policy and financing mechanisms into the same model as provincial coordination, provider execution and electricity. Its potential to become a distinctive systems approach is presented as a hypothesis, with explicit mechanisms, limits and outcome tests.
Explore the revised architecture →01 / THE NATIONAL LAYER
A common framework above local resources
A national framework cannot coordinate resources that are not first visible and usable locally. So the argument now moves from the national layer down into provincial implementation.
MIIT’s September account reports 17 approved regional interconnection nodes, more than 70 transmission links and an intelligent-compute estate of 2,185 EFLOPS (FP16, as specified in the Digital China conference account). It also describes national monitoring and platform connections across 31 provincial-level regions. These figures describe different things: approved nodes, network reach and installed compute are not a measure of nationally movable workloads. MIIT account, 14 September.
Latest scale check: National Data Administration figures for end-July put total intelligent compute at 2.45 million PFLOPS (FP16), with 1.45 million PFLOPS — about 60% — inside the national monitoring scope. This is a visibility measure, not a schedulable-capacity measure. National Data Administration, 29 August.
Shared identifiers, interfaces and visibility
National service functions and the 1+M+N node programme provide a framework for discovery and coordination. Seventeen regional nodes were approved; this is not a count of fully operational national schedulers.
Provincial and operator platforms
Connect existing schedulers, operator resources and eligible local data centres. Local permissions and operating controls still matter.
Cross-region resource access
Discover and request capacity beyond the province. Xiong’an provides a platform example; Fujian–Ningxia provides a specific workload-transfer pilot.
Resource pools, clouds and data centres
Workloads execute on provider-managed systems. Chip compatibility, latency, data location and customer contracts determine what can actually move.
The diagram is our synthesis of the evidence, not an official command hierarchy. It does not imply that one national scheduler dispatches every task or that every provincial platform is the same system.
02 / INTRA-PROVINCIAL COORDINATION
First connect the resources inside the province
Provincial aggregation is necessary but not sufficient. The next step is whether those resources can participate in networks that cross administrative boundaries.
Hebei’s regional-node implementation prioritises existing scheduling platforms and telecom resources, then progressively brings provincial data centres and eligible government, university and enterprise resources into managed pools. Sensitive resources are subject to separate controls. This is an aggregation programme with staged onboarding, not evidence that all local compute is freely interchangeable. Hebei government-hosted implementation account, 21 September.
Guizhou offers a complementary example: its regional-node consortium includes the three telecom operators and Guizhou’s grid company. That establishes institutional coordination between compute, networks and electricity; achieved dispatch volumes and returns remain separate questions. MIIT / Guizhou Communications Administration, 4 September.
Keep the Hebei platforms distinct: GDS reports participating in the launch of Hebei’s provincial monitoring/scheduling platform. The Hebei regional interconnection node and Xiong’an scheduling centre are separately described initiatives. Participation in one does not prove ownership or control of the others. GDS disclosure.
03 / INTER-PROVINCIAL COORDINATION
Xiong’an connects a provincial base to a wider resource network
Cross-region access changes the geography of compute. The Fujian–Ningxia case adds the energy dimension: moving a workload can also move electricity demand.
The Xiong’an authority announced the integrated scheduling centre’s launch on 12 September. Its platform connects 44 resource pools across 19 provincial-level regions and covers all eight national compute hubs. That is a concrete example of computing resources being connected across regions; the announcement does not quantify recurring workload movements, completed transactions or returns. Xiong’an authority, 13 September.
A related model-service platform offers model access through a common interface. Token or model-request routing is a different function from migrating a running training job between data centres. Its existence should not be read as evidence that all underlying GPUs, models and workloads are interchangeable. Xiong’an platform account, 18 September.
04 / OPERATIONAL PILOT · EVENT 16 SEPTEMBER · SOURCE 22 SEPTEMBER
Fujian → Ningxia: moving compute changes where electricity is used
One pilot does not prove nationwide fungibility. Before drawing that conclusion, the next section separates the capabilities that are often bundled together in policy language.
State Grid’s account on SASAC describes pilot workloads moving from Fuzhou and Xiamen to Zhongwei through China Mobile’s compute-network scheduling platform. Fujian’s load fell while Ningxia’s rose to absorb surplus midday solar. Power-load management, virtual-power-plant and compute-network systems exchanged data. State Grid / SASAC source.
What changes in the thesis: this is an operator-reported execution example of workload placement responding to energy conditions. It advances the evidence beyond infrastructure co-location and general policy intent.
| Step | What the link means |
|---|---|
| Workload → scheduler → geography | An eligible task is placed at a different location. |
| Geography → data centre → electricity demand | The receiving site consumes power to run the work; demand shifts between locations. |
| Electricity demand ↔ renewable availability | The receiving load can align with available generation. Operational signals and results inform the next decision. |
25 September evidence extension: Shanghai and Zhejiang now provide additional bounded examples of workloads responding to electricity conditions. Shanghai’s official report includes an electricity-price-aware cross-region inference test; Zhejiang’s report describes partial Jiaxing-to-Qinghai migration. This strengthens the mechanism while leaving national scale, portability across heterogeneous stacks and net commercial economics unresolved.
Pilot evidence, not a national commercial result. The cited account does not establish MW/MWh scale, workload class, migration cost, service-level impact or recurring profitability. No VNET, CATL, KC or GDS participation is inferred.
05 / EVIDENCE DISCIPLINE
Four capabilities that should not be conflated
Only after separating these capabilities can we ask the commercial question: who operates them, who pays, and whether coordination improves utilisation or returns.
| Capability | What it does | What it does not establish |
|---|---|---|
| Monitoring | Counts and observes resources and operating conditions. | Permission to dispatch workloads, or cross-provider compatibility. |
| Marketplace | Lists products and connects supply with potential customers. | Completed transactions, utilisation or customer retention. |
| Workload scheduling | Places eligible work on selected resources under constraints. | Seamless migration of every application or lower total cost. |
| Token / model routing | Directs an inference request through a model-service interface. | Moving the underlying model, training job or physical compute estate. |
National marketplace counts are evidence of platform reach: MIIT reports over 10,000 registered enterprise users, over 2,000 listed products and over 300 models. They are not active-paying-user or transaction-volume metrics. Source and definitions.
06 / HOW THE THESIS CHANGES
Coordination becomes a delivery and economics question
System inference: value may depend on connecting usable compute, location and energy conditions, as well as building capacity. We now have distinct evidence for national visibility, provincial aggregation, cross-region platform reach and one compute–electricity pilot. They should be tested separately.
- GDS: direct participation evidence exists for the Hebei platform launch; this is not an incremental customer award or proof of scheduler revenue.
- VNET–CATL: energy coordination is relevant to the partnership thesis, but the pilot establishes no involvement, funding or revenue for either company.
- KC: workload orchestration is conceptually relevant; these sources do not identify KC as a national or provincial scheduler.
- Alibaba: the 20 GW ambition raises the importance of power and location. It does not turn the national platforms into Alibaba procurement commitments.
Next tests: recurring dispatched workloads; service levels; measured energy and migration costs; payment and settlement; and the share of savings retained by operators. Valuation assumptions are unchanged.
Power & energy interpretation → · Control-plane interpretation → · Dashboard →
Sources and context
Event dates and publication dates are kept separate. Official announcements and company reports are attributed; performance and economic outcomes require their own evidence.
- MIIT / Chongqing Communications Administration · national monitoring and network figures 14 September 2026; sector figures refer to end-June.
- MIIT / Guizhou Communications Administration · regional-node consortium 4 September 2026.
- Hebei cyberspace authority · provincial implementation account 21 September 2026; republishes Hebei Daily reporting.
- Xiong’an authority · scheduling-centre launch and resource coverage 13 September 2026; event 12 September.
- Xiong’an official website · model-service and scheduling platforms 18 September 2026; republishes Hebei Daily reporting.
- SASAC / State Grid · Fujian–Ningxia compute–electricity pilot 22 September 2026; event 16 September.
- GDS · Hebei platform participation and Langfang projects 14 September 2026; company disclosure.
- National Data Administration · national monitoring trial platform3 September 2026; 1,462 facilities and 1.45 million PFLOPS monitored.
- CCTV via Digital China · resource islands and Kaifeng network adaptation24 September 2026.
- State Grid Shanghai · coordinated data-centre / compute–electricity test15 July 2026; underlying exercise in July.
- China Energy News · Zhejiang token-level energy/carbon and workload migration test31 August 2026; reports Jiaxing-to-Qinghai partial task migration.