Architecture evidence · 17 October 2026
Computing power depends on electrical power—and on where AI can be delivered
China's 16 October State Council discussion of AI Plus again placed suàn diàn xiétóng (算电协同), computing–electricity coordination, alongside advances in models, algorithms and data. The strategic ambition is clearer than the commercial outcome: a data centre becomes useful infrastructure only when electricity, network connectivity, installed hardware and paying workloads arrive together. The State Council account reinforces the policy direction; it is not proof of nationwide operational integration.
Ulanqab makes the constraint tangible. In September, VNET's local chief engineer described the Dongying 500kV project as a needed step for its campus development. The Xinhua Finance account distinguishes an operating first phase from future development-scale capacity. It also raises a financing question: grid upgrades, campus electrical systems, customer hardware and energy storage need not sit on the same balance sheet. The cost and ownership of each determine who ultimately captures the economic value.
China's proposed 16 October AI technology-export revisions add a separate geographical question. The ministry's accompanying explanation distinguishes transfers of controlled technology from supplying model access through APIs and cloud services, while outlining facilitated arrangements for some Chinese-operated overseas computing facilities. This may preserve multiple routes to international demand, but it does not establish that more workloads will be hosted domestically. The final rules, customer deployment choices and actual cloud contracts remain the tests.
Evidence integration · 17 October 2026
A domestic accelerator is moving into a named telecom computing campus
The newly reported Ascend 950DT project at China Telecom’s Inner Mongolia Cloud Computing Information Park in Horinger brings together several elements of the emerging architecture: Huawei’s accelerator and high-bandwidth interconnect, a telecom operator’s computing campus, and a three-dimensional design that separates cooling, computing and power infrastructure. Reporting on 14 October describes a first 10,000-accelerator project, with provision for a 30,000-card cluster, while Huawei’s northern 3D data-centre structure has reached its topping-out milestone. These are project and construction claims, not evidence that 10,000 accelerators are already delivering paid computing services.
The reported design includes 1,024-card supernodes and Huawei’s Lingqu 2.0 interconnect. Huawei had previously positioned the 950DT for training and the decode stage of inference, where memory bandwidth and interconnect performance are particularly consequential. The architectural link is tangible: chips, networking, liquid cooling and power distribution are being planned together at a named site. The economic link remains unverified until installed hardware, commissioning, utilisation and customer workloads are disclosed.
Sources: 14 October regional project report; Huawei’s Ascend 950DT architecture and roadmap.
Architecture evidence · 16 October 2026
Compute–energy integration has to become a commercial contract
China's proposed compute–energy systems are developing beyond the simple pairing of a data centre with renewable electricity. At Beijing's 15 October forum, speakers emphasised flexible load, storage, electricity-market participation and the operational details of measurement and compensation. Those details determine whether workload scheduling actually produces a bankable saving or new service income.
Yunding Technology's 14 October memorandum with Huawei adds an industrial-demand example: existing energy-sector applications, planned Ascend compute, and a proposed token-service offering. The link is confirmed strategic cooperation, not evidence of an installed cluster or profitable token sales. Sources and evidence boundaries.
Recent demonstrations illustrate why the connections between layers matter. China Telecom and Huawei have reported training across three separated computing centres, while Huawei and China Mobile Hubei have tested storage-managed KV caching to improve long-context inference throughput. The results are promising but vendor-reported. They show possible gains from networking and software, not yet the economics of a nationally operated system.
The commercial layer between compute and customers
The emerging architecture is not only chips, networks, data centres and power. Model selection, workload routing, usage metering and billing determine whether those assets become usable services. Shanghai’s Haina launch makes that commercial layer easier to see, without proving national-scale execution.
1 OCTOBER 2026 · ARCHITECTURE UPDATE
The system is becoming more capable domestically — but it still has an offshore pressure-release valve.
Current-state change: DeepSeek’s Ascend-native infrastructure shows software adaptation reaching important model kernels and communication layers. At the same time, Tencent is reported to be leasing advanced AI compute from Oracle in Southeast Asia, showing that Chinese demand can route around domestic hardware constraints by moving workloads offshore.
Boundary: this does not weaken the case for domestic coordination; it clarifies it. China is simultaneously building domestic substitutes and using external capacity where available. The relevant test is therefore how much workload can be served efficiently onshore, not whether every Chinese workload remains inside China.
See the 1 October evidence and limits →30 SEPTEMBER 2026 · ARCHITECTURE UPDATE
Policy finance, measurable AI output and compute–energy design are beginning to connect.
Current-state change: the computing-power network is now explicitly inside PSL support; Beijing has piloted Token-linked credit underwriting; Baotou has begun building a project that couples grid-side storage with distributed intelligent compute and scheduling; and a 1,024-card Huawei SuperPoD has moved into a major operator environment.
Boundary: these developments strengthen different layers of the architecture, but they do not establish one integrated national commercial system. Funding eligibility is not a project drawdown; Token-linked credit is a local pilot; Baotou is under construction; and the SuperPoD deployment still needs utilisation and economics evidence.
See the 30 September evidence and limits →ARCHITECTURE UPDATE · 25–26 SEPTEMBER 2026
Power signals, network reachability and service-grade output now sit more clearly inside the same operating loop.
Current-state change: Shanghai documents electricity-price-aware cross-region inference placement; national monitoring shows substantial resource visibility; Kaifeng and CAICT transport evidence show that low-latency paths, task-based scheduling and compute-network intelligence are still required to turn visibility into usable remote capacity. The 26 September review also adds an emerging industry measurement layer around effective tokens, service quality and cost per useful output.
Boundary: these are different maturity states. A resource can be visible without being reachable, reachable without being portable, and portable without being economically attractive. The architecture therefore remains federated and partially connected rather than one fungible national pool.
22 SEPTEMBER 2026 · THESIS UPDATE
Policy, finance and coordination now sit inside the architecture
The revised model connects national policy and financing mechanisms with provincial aggregation, inter-provincial access, provider execution and compute–electricity feedback. It explicitly tests how this combination could bridge supply constraints and develop into a distinctive national systems approach.
Interpretation: this strengthens the case for coordination above individual providers. Monitoring coverage, platform access and actual execution remain different evidence states.
Read the evidence and implications →Financing and compute–energy co-design now sit more clearly inside the same national systems architecture.
The newest evidence adds a policy-finance route for the computing-power network, a local Token-linked credit mechanism, construction-stage storage–compute co-design and a major operator deployment of a 1,024-card domestic SuperPoD. Together they extend the architecture beyond connectivity alone.
Standards, monitoring platforms and operating tests now support different parts of the same architecture: resource visibility, cross-region access, workload placement and compute–electricity feedback. They remain different capabilities and should not be collapsed into one claim.
Still unproven: routine national workload portability, uniform heterogeneous scheduling, commercial settlement and superior total economics.
A potentially different way to build and operate AI infrastructure.
Now that the mechanism is visible, the next question is implementation: what policy, financing and delivery structures could make it real?
Our thesis is that China may be developing a distinctive national systems approach: linking hardware, software, data centres and electricity through coordinated government policy, financing mechanisms and resource coordination within and between provinces. Its potential novelty lies in how these elements work together at national scale. The evidence does not yet establish a wholly new model or superior results.
The architectural question is whether this combination can bridge some supply constraints by making more of the available infrastructure usable, accessible and productive. That means examining both new capacity and the practical ability to place suitable workloads where compatible compute, power and network capacity are available.
Coordinate the conditions for a national system
National priorities, standards and infrastructure planning shape deployment. Policy-backed financing mechanisms and commercial investment support delivery; project-level commitments and disbursements must be evidenced separately.
Make distributed resources visible and accessible
Identifiers, monitoring, interconnection and service rules can connect different operators and regions. Monitoring, marketplaces, model routing and workload scheduling have different capabilities and maturity.
Connect local resources and demand
Provincial and operator platforms aggregate eligible capacity and connect existing systems, subject to permissions, contracts and technical compatibility.
Match eligible work to a wider supply base
Cross-region access can connect demand with suitable resource pools. Placement must account for latency, data movement, energy conditions and service commitments.
Run the workload on an appropriate local system
Cloud software, compilers and runtimes operate compatible compute and memory pools. Cluster networks, cooling, data centres and electricity determine their usable performance. Tight coupling belongs within suitable local clusters; long-distance links support resource access and workload placement.
Coordinate where and when flexible demand runs
Energy availability and price can inform workload placement; executing the work changes local electricity demand. Shanghai now provides an official bounded operating example, while Fujian–Ningxia and Zhejiang add separate workload-shifting evidence. Repeatability, workload eligibility and net economics remain to be established.
More useful AI output from the resources available
Test utilisation, completed work, total cost, reliability, energy use and deployment speed. Installed FLOPS, connected capacity and policy targets alone cannot establish this outcome.
This is our synthesis, not an official command hierarchy. National hubs, regional interconnection nodes, provincial platforms and operator schedulers are distinct structures whose roles can overlap. It does not assume one national scheduler, interchangeable chips or a single nationwide supercomputer.
Supporting evidence: national interfaces · provincial aggregation · cross-region access · compute–electricity operating evidence.
How the architecture could bridge supply constraints
These are proposed mechanisms to test, not quantified gains. Coordination may reduce fragmentation and improve access to capacity; physical shortages and software limits still require direct solutions.
| Constraint | Potential system response | Remaining limit / test |
|---|---|---|
| Scarce high-performance accelerators | Match each workload to compatible hardware; improve utilisation and use alternative resource pools where software permits. | Scheduling cannot manufacture chips or guarantee equivalent performance. Measure completed useful work and portability costs. |
| Memory capacity and bandwidth | Workload-aware memory and storage tiers may relieve some capacity pressure within suitable systems. | Remote capacity and storage do not replace local HBM bandwidth. Measure end-to-end performance. |
| Capacity separated from demand | Provincial aggregation and cross-region access may connect spare resources with eligible customers. | Discovery is insufficient without compatibility, data permissions, service levels and completed transactions. |
| Power and location mismatch | Place flexible work where power is available; coordinate timing with energy conditions and local infrastructure. | Grid connections, cooling, latency and migration costs remain binding. Measure net energy and service outcomes. |
| Capital and fragmented build-out | Align infrastructure planning, standards and financing mechanisms with usable demand and network connections. | Policy support is not committed funding or viable demand. Track disbursements, commissioning and sustained utilisation. |
What would substantiate the thesis? Recurring cross-provider or cross-region execution with measured improvements after networking, software, migration and financing costs. Low utilisation, persistent incompatibility or dependence on support without durable demand would weaken it.
National compute coordination is becoming an explicit architecture layer New · 21 Sep
21 September 2026 update. Newly registered National Integrated Computing Power Network standards materially strengthen the control-plane thesis. The programme now spans compute-grid connection, monitoring interfaces, resource identification, scheduling, multidimensional billing, operations/matching transactions, compute–electricity coordination and assessment of data-centre adjustable-load potential. These are draft/registered technical projects rather than proof that one national scheduler already controls workloads, but they move the thesis from broad policy language toward defined interfaces and operating rules.
Updated abstraction: enterprise/cloud control planes sit below an emerging national coordination layer that can make heterogeneous compute resources increasingly identifiable, monitorable, schedulable, billable and tradable across regions and operators. The physical AIDC and energy layers therefore connect upward to software orchestration and sideways to the power system.
Primary evidence: National standards register — adjustable data-centre load potential · National standards register — related integrated-compute-network projects
ARCHITECTURE UPDATE · 23 SEPTEMBER 2026
Agent workloads strengthen the case for system-level optimisation
Current-state change: the architecture now explicitly connects workload characteristics to memory/KV cache, interconnect, scheduling, network geography and power. This is the architectural result of the latest evidence—not a duplicate of the evidence record.
The companies reveal how the system is taking shape.
Select any layer to see the current interpretation. The economic endpoint is not installed FLOPS or MW in isolation, but useful AI output relative to capital, power and operating complexity.
Coordinated policy and finance are part of the architecture.
This is the delivery layer. Read it as evidence about whether the architecture can move from national direction to funded, operating infrastructure.
The Policy Hub now separates four questions: what Beijing intends, the mechanisms chosen to pursue it, what has actually been implemented, and whether measurable outcomes follow. This lets policy strengthen or weaken an architectural hypothesis without ever becoming a shortcut from policy to stock.
National direction, financing mechanisms, project-level delivery
The December 2023 five-department implementation opinion combines government coordination with market mechanisms. It supports eligible green data-centre, transmission-network, urban-compute-network and compute–electricity projects exploring infrastructure REITs, alongside stronger policy-based finance support. This gives financing a defined role in the framework; it does not establish a single national funding pot or an award to any named company.
NDRC · national integrated computing network implementation opinion · 25 December 2023.
Our interpretation: the potentially distinctive feature is the attempt to align technical coordination with national planning and financing. Whether that alignment produces better economic outcomes remains an empirical question.
AI+ sets the umbrella direction.
The State Council's 2025 AI+ opinion links domestic AI-chip innovation and software ecosystems with ultra-large intelligent-compute clusters, the integrated national compute network, East Data–West Computing and greater coordination of data, compute, electricity and networks. This is strategic direction rather than evidence that every layer is already operational.
Policy convergence matrix
| Programme | Chips | Software | Memory | Network | Compute | AIDC | Geography | Energy |
|---|---|---|---|---|---|---|---|---|
| AI+ | ||||||||
| Compute Interconnection | ||||||||
| 1+M+N nodes | ||||||||
| AI + ICT 2026–28 | ||||||||
| AI + Energy |
Dots show material policy intersection, not funding, company selection or proof of deployment.
Implementation ladder
What central policy says China wants to achieve.
Specific ministries convert direction into programmes and targets.
Interoperability, SuperPods, heterogeneous compute and networking become defined categories.
1+M+N regional/industry nodes and other programmes move into build-out.
Utilisation, scheduling, cost, energy and useful output determine whether policy actually worked.