When Compute Becomes a Commodity: The Rise of China’s Compute-Scheduling Platforms and How to Vet Them

For most of the past decade, buying compute in China meant signing a multi-year contract with one of a handful of state-backed IDC operators or hyperscale cloud providers, wiring a large upfront deposit, and committing to a fixed rack or GPU quota you might or might not actually use in month two. The model was built for large enterprises; it was not built for startups, research teams, or overseas firms that need a few hundred card-hours of H800-class training, a batch of inference jobs, or a week-long simulation run.

That model is now being formally dismantled from the top down. In April 2026, the Ministry of Industry and Information Technology (MIIT) issued the Special Action to Empower SMEs with Inclusive Compute (工信厅通信〔2026〕14号), a policy document that for the first time explicitly instructed the industry to develop two novel business formats: “compute banks” (算力银行) and “compute supermarkets” (算力超市). The goal, articulated in the 15th Five-Year Plan logic, is to turn compute from a heavy-asset, long-commitment input into a light-service, on-tap utility — what officials routinely describe as “using compute like water and electricity.”

The scale of activity behind this policy push is already hard to ignore. According to MIIT and CAICT data cited at the 2026 Big Data Expo in Guiyang in May:

• Daily national Token invocations reached 140 trillion by March 2026 — up from roughly 100 billion in early 2024, a more than 1,000× increase in two years.
• The national unified compute-network monitoring, dispatch and experimental-verification platform launched by the National Data Administration in August 2025 had integrated 1,462 compute facilities by September 2026, with over 60% of national compute capacity under unified monitoring.
• More than 20 dispatch entities — including the three state telecoms, national research bodies, and regional operators — formally joined the national platform at the 2026 Big Data Expo, creating the institutional backbone for cross-region trading.

2. The “bank” and the “supermarket”: what they actually are

The two concepts are often lumped together in English reporting, but they serve distinct economic functions. Guo Liang, chief engineer of the Cloud Computing and Big Data Research Institute at the China Academy of Information and Communications Technology (CAICT) — the technical body that drafted MIIT’s action plan — gave the most widely-cited framing in a May 2026 interview with China News Service:

The analogy is useful because it flags what is genuinely new. In 2023 and 2024, cloud providers already sold GPU instances by the hour, but the compute you bought came out of that provider’s own data centers and was metered on that provider’s own bill. What is new in 2026 is the multi-seller, cross-region, federated dimension:

Regional variants have proliferated quickly. Shanghai Telecom’s “AI STORE · Zhiyun Shanghai” links the Qingpu and Lingang hub clusters and supports WeChat self-checkout for individual buyers alongside custom enterprise orders; Ningxia — the first mover — launched a province-level integrated compute trading and dispatch platform several years ago, which remains the western anchor of east-data-west-compute trading; Shandong, Zhejiang, and several other provinces have all announced supermarket-style pilots in 2026. On May 9, the Henan Airport Intelligent Computing Center (built with Chinese chipmaker Teco-Epoch/太初元碁 hardware) became the first central-China hub to offer Token-billed DeepSeek access with no upfront hardware commitment.

3. Four risks the platform model introduces

For an overseas buyer — whether you are an AI startup running fine-tuning jobs, a hedge fund needing batch inference, a life-sciences team running molecular simulations, or a consulting firm validating a Chinese partner’s claims — this is, on paper, great news. You no longer need a Chinese legal entity, a three-year contract, or a relationship with a specific IDC. A few clicks, a wire transfer in digital RMB or cross-border settlement, and you have compute.

But convenience masks structural risk. The same characteristics that make platforms attractive — low commitment, multi-seller inventory, Token-level granularity — also make them harder to vet than a traditional single-operator contract. Four risks stand out:

4. Who runs these platforms? Mapping the five operator types

Not every “compute supermarket” is created equal, and the operator category tells you a great deal about the due-diligence questions you should be asking. Based on publicly disclosed launches in 2026, platforms currently fall into five archetypes:

The Fujian case is illustrative. In April 2026, at the 9th Digital China Summit, the Fujian Integrated Compute Resource Public Service Platform went live, built and operated by Start Group (实达集团, 600734.SH) — a listed company controlled by Fujian Big Data Group, the provincial SOE. The platform implements an integrated “compute + model + metering” stack, runs a “compute mall” for GPU models and Token billing, and cross-dispatches to partner hubs in Ningxia and Xinjiang. For a buyer, this is a structurally different proposition from a well-designed website run by a 50-person startup: the licensing is on a state-owned operator of record, the data-location story is auditable, and recourse channels exist.

5. A seven-step verification playbook before you transfer any funds

If you are considering buying compute from a Chinese scheduling platform — whether for a single fine-tuning run or a multi-month inference contract — the following seven checks, performed against official Chinese government records before signature, will filter out the vast majority of platform-washing and counterparty risk.

1. Pull the operator’s official business registration record (工商登记档案) from the State Administration for Market Regulation (SAMR) via the National Enterprise Credit Information Publicity System (国家企业信用信息公示系统). Confirm: (a) the registered Chinese company name exactly matches the contracting entity on your PO; (b) business scope (经营范围) explicitly lists compute services, cloud computing, IDC, or internet information services — an operator whose scope is “software development” or “tech consulting” is not licensed to resell compute; (c) registered capital, paid-in capital, and establishment date are consistent with the scale of the platform they claim to run.
2. Verify telecom-licensing coverage end-to-end. Cross-check the operator (and every subsidiary that will deliver services to you) against the Ministry of Industry and Information Technology’s public license database. At minimum you need to see a valid B11 IDC license for the province(s) where your workload physically runs, plus B25 ICP if you are consuming via API/web, and B21 EDI if third-party sellers settle transactions through the platform. Confirm the license is not in revoked, suspended, or “rectification” status.
3. Trace the equity chain to the actual controller. A platform branded as a provincial “government” marketplace is sometimes operated by a mixed-ownership subsidiary several layers down. Use official AIC records to follow ownership to the natural person or ultimate state entity, and cross-check against sanctions lists, export-control entity lists, and your internal related-party registers.
4. Check the administrative-penalty, abnormal-operation, and equity-freeze records. The National Enterprise Credit Information Publicity System records these directly. Prior fines for unlicensed telecom operations, energy-consumption violations at an affiliated data center, tax arrears, or abnormal registered address (经营异常) should be treated as bright-line red flags — they can foreshadow forced shutdowns mid-contract.
5. Map the intellectual-property stack. Search the China National Intellectual Property Administration (CNIPA) database for patents, software copyrights (软件著作权), and trademarks filed by the operator. A genuine scheduling-platform operator — as opposed to a reseller — should hold core software copyrights for its dispatch, metering, or billing systems, and ideally participate in industry standards. Absence of any registered IP is a strong signal of a re-branded aggregator. Our intellectual-property verification service can consolidate these filings with English translations within a few working days.
6. Demand auditable evidence of underlying capacity. Before signing a large prepaid agreement, ask for: (a) the name and location of the physical data center(s) that will host your workloads; (b) the corresponding PUE test reports and energy-quota permits; (c) a sample of real-time monitoring data showing GPU utilization over the prior 30 days (to detect oversubscription); (d) the chip SKU, not just a “equivalent” marketing label. Reject answers that cite only “national hubs” or “partner resources” without naming facilities.
7. Pull the operator’s bidding and government-procurement record. For platforms that market themselves as serving government or state-owned clients, verify those claims against official public-tender platforms (中国政府采购网, 全国公共资源交易平台). The existence and size of awarded government contracts is a practical proxy for both license compliance and operational maturity — contracts with provincial big-data bureaus or telecom operators would not be awarded to an unlicensed shell.

6. How ChinaBizInsight can help

The compute market China is building is genuinely novel — a hybrid of grid economics, cloud marketplace, and industrial policy, with participation rules still being written in real time. For overseas buyers, that means the surface area of counterparty risk is broader than in traditional IDC procurement: you are not just diligencing a data center, you are diligencing a trading platform, its upstream suppliers, its settlement layer, and its licensing status all at once.

This is precisely the kind of multi-layered Chinese-counterparty verification that ChinaBizInsight specializes in. Starting from the operator’s Chinese-registered entity name, we can deliver an Official Enterprise Credit Report drawn directly from SAMR / National Enterprise Credit Information Publicity System records — giving you the registration, business scope, penalty history, and equity structure in English, with official chops. For larger commitments, our Professional Enterprise Credit Report layers in beneficial-owner tracing, affiliated-entity networks, IP portfolios, litigation records, and government-procurement history so that you see who actually runs the platform before you wire any prepayment.