The “8+10+3” Map Decoded: How Overseas Companies Can Use China’s Regional Compute Clusters to Find the Right Partner

1. What “8+10+3” Actually Means

When Chinese officials refer to the “8+10+3” spatial layout of the national compute network, they are not describing a marketing slogan. They are describing the legally defined zoning of where data centers are allowed to be built at scale, where cheap green electricity is available, and where inter-city optical backbones have been pre-wired. A company registered inside one of these zones faces an entirely different cost curve, regulatory path and customer mix than one registered a few hundred kilometers outside it.

The numbers break down like this, per the official interpretation from the National Development and Reform Commission (NDRC) and the National Data Administration (NDA):

• “8” — eight National Compute Hub Nodes: four in the east (Beijing-Tianjin-Hebei, Yangtze River Delta, Greater Bay Area, Chengdu-Chongqing) and four in the west (Inner Mongolia, Guizhou, Gansu, Ningxia).
• “10” — ten National Data Center Clusters physically carved out within those eight hubs. These are the only parcels of land where very large (>100 MW) data center projects routinely receive power, land and bandwidth approvals without being rejected as “redundant construction.”
• “3” — three Compute-Power-Electricity Synergy Zones, the newest layer added during 2025–2026, starting with Xinjiang as the national pilot. These zones are wired directly to wind and solar farms via dedicated green-power direct-supply lines and are meant to push electricity costs for AI workloads well below the RMB 0.30/kWh threshold.

2. The Big Divide: East Carries the Workload, West Carries the Watts

For four years the “East Data West Computing” program has been pushing compute capacity westward. But by 2026 the industry has settled into a pragmatic two-tier model — and understanding it is the fastest way to evaluate a Chinese partner’s claims.

The economic logic is brutal. A typical 10,000-P training cluster consumes ~35 million kWh per year. At RMB 0.80/kWh in the Yangtze River Delta that is RMB 28 million (~USD 3.9 million) in electricity alone. At RMB 0.30/kWh in Qingyang or Zhongwei, the same workload costs RMB 10.5 million (~USD 1.5 million). That single line item — before GPUs, before staff, before bandwidth — is a 60%+ swing. It explains why every major Chinese model developer now runs the bulk of pretraining west of the Heihe-Tengchong line.

But physics does not move. Light travels ~200 km per millisecond through fiber. Chengdu to Shanghai is 2,000 km straight-line — a 10 ms one-way trip at the speed of light, ~16–20 ms in a real routed network, before you add any GPU processing. That is why roughly 70% of global AI compute load in 2026 is inference, not training, and why latency-sensitive inference will not move west any time soon — no matter how cheap the kilowatt-hour.

3. Reading the Map by Workload: Matching Your Need to the Right Zone

Location is not a proxy for quality — but it is a fast first-pass filter. If you know what you are buying, the registered address of your Chinese counterparty tells you whether they are structurally positioned to deliver it.

4. Three Signals a Registered Address Is Telling You

The address on a Chinese company’s business license is not a formality. It is one of the most information-dense fields in the entire AIC record. Here are the three things every overseas buyer should extract from it before even opening a pitch deck.

All three signals — the specific cluster address, the ownership type, and the change history — are available directly from the Official Enterprise Credit Report pulled from the National Enterprise Credit Information Publicity System, and can be cross-checked against power-quota approvals and land-use certificates in a deeper customized credit report.

5. Cluster-by-Cluster Cheat Sheet

The ten national data center clusters are not interchangeable. Here is the working map we use internally when advising overseas clients on counterparty screening.

Beyond the ten, Xinjiang (Hami Yizhou, Karamay) is the new frontier in 2026. The first multi-user green direct-supply projects (one wind+solar farm serving multiple data centers over private lines) were approved in May 2026; delivered tariffs can fall below RMB 0.31/kWh, and the Xinjiang-to-Chongqing green-power trunk is already live at 16.2 ms one-way with auto OEM customers including Changan and Seres. Xinjiang is not yet part of the “8+10” backbone, but it is the strategic “3” and worth watching for long-horizon training workloads.

6. Seven-Step Location-Based Diligence Playbook

Before signing a multi-year compute or data-center contract with a Chinese counterparty, run these seven location-focused checks in sequence. Steps 1–3 come from public government registries; steps 4–7 typically require a structured due-diligence report.

1. Pull the Official Enterprise Credit Report and confirm the registered address, unified social credit code, enterprise type (SOE / private / listed subsidiary / foreign-invested), and the precise business scope. Flag any company whose scope does not explicitly include data-center services, IDC services, compute services or internet information services but is pitching you a compute product.
2. Map the address to the hub/cluster polygon. Confirm whether the registered (and any disclosed operating) address falls inside one of the 10 clusters or 3 synergy zones. If it does not, ask the counterparty to disclose the actual data-center premise by full street address, and re-run the check on that entity (often a subsidiary or SPV).
3. Verify the telecom license(s). Cross-check whether the operating entity (not a parent or affiliate) holds a valid B11 IDC, B25 ICP or B21 EDI value-added telecom license issued by MIIT or the provincial regulator, registered to the same province as the operating site. License-less resale is common and creates Section 232 / BIS exposure for foreign customers under certain contract structures.
4. Pull outbound-investment records. Confirm whether the company actually owns, rather than resells, capacity in the hubs it claims — particularly in western synergy zones. A Shanghai marketing entity with no equity in any Gansu/Ningxia/Xinjiang operating subsidiary is a broker, not a provider.
5. Check for address-churn, penalties and tax anomalies (administrative penalties, business-irregularity listings, tax-violation records). Companies that have repeatedly moved their registered address within short windows, or that carry environmental penalties for exceeding energy caps, are at structural risk of having their power allocation cut — a death sentence for an IDC.
6. Request evidence of power quota and PUE compliance (energy allocation certificates, completion acceptance documents, third-party PUE test reports) for the specific campus. A press release announcing a “500 MW AI park” means nothing until the local DRC has allocated the corresponding energy quota, the grid-connection document is signed, and a third party has tested PUE.
7. Run a small, paid pilot workload on the exact cluster you are being pitched, measuring round-trip latency, jitter, and token/s throughput from your own geography, before committing to any prepaid capacity commitment. Geography on paper ≠ network performance in production.