How Qingyang Is Building an AI Compute Utility From Green Power

Qingyang’s transformation is not principally a story about building more data-center space. It is an attempt to combine electricity, networks, AI accelerators and workload scheduling into a metered computing service.

In August 2026, Qingyang reported 185,500P of cluster capacity, with intelligent computing accounting for more than 99% of the total. Local reporting also said the cluster supported more than 25% of China’s independent-model AI token demand. Those are reported operational figures, not independently audited market-share estimates; later August reporting placed total capacity at 215,000P, underscoring how quickly the build-out was changing. 3538

The central advantage: low-cost green electricity

AI data centers are electricity-intensive, so Qingyang’s strategy starts with power rather than real estate. The China Mobile park reported a green-power aggregation direct-supply arrangement supported by a planned 2 GW wind-and-solar base, with green electricity exceeding 80% of supply and a delivered price of 0.398 yuan per kWh. The company said this was roughly 0.2–0.3 yuan per kWh cheaper than eastern locations. 39

The wider model links renewable generation, the grid, data-center load and energy storage—known as source-grid-load-storage integration. A National Data Administration case study describes a wind-and-solar project paired with a green-computing operations platform that coordinates supply and demand, aiming to provide more than 90% renewable electricity to the park through a virtual dedicated-power arrangement. 52

That integration matters because a GPU cluster needs dependable power around the clock, while wind and solar output varies. The system does not eliminate that constraint, but it is meant to make locally generated renewable power more usable for continuous computing loads.

Fast, high-density capacity deployment

Qingyang’s construction approach is designed to add capacity in stages without redesigning the whole campus each time. China Mobile reported completing a 140-mu site plan, including 300 MW of power capacity and site utilities, then building and commissioning data-center buildings in phases. 46

The Zhihui Blueprint Western Intelligent Computing Center shows the scale of the next wave. Its two phases are planned to provide 136,000P in total. Phase one is planned with three data-center buildings, a power center, a 110-kV substation and 8,860 high-density cabinets, for 108,000P of capacity. 37

Qingyang’s official data-center construction guidance encourages prefabricated, modular and assembled construction, alongside integrated energy systems. 57 However, the available sources do not substantiate the claim that the park itself was built on former cornfields, so that detail should not be treated as established fact.

Why eastern customers can use a western hub

Moving AI workloads west only works if network performance remains acceptable. Qingyang reports a full-optical “1–3–7–14 ms” latency circle: latency to Xi’an and Lanzhou is reported at under 3 ms, while the Greater Bay Area is about 14 ms. 12

That does not make every latency-sensitive application location-agnostic. But it can make remote inference, model fine-tuning, batch processing and other suitable workloads more feasible to schedule into Qingyang rather than keeping all compute next to eastern demand centers.

An ecosystem built around chips, compute, models and applications

Physical capacity does not create demand on its own. Qingyang’s ecosystem includes more than 700 firms across the chip–compute–model–application chain, according to China News Service reporting. The reported participants span chip suppliers, compute-service providers and model/application companies. 2

The city has also used coordinated enterprise recruitment. By late 2025, local authorities said they had contacted 6,463 digital-economy firms, signed agreements with 1,522 and brought more than 500 firms to the city. 3 These figures are official reporting, but they illustrate the strategy: use early infrastructure and anchor tenants to make the next deployment easier.

Local applications help anchor utilization. Reported examples include energy exploration, green agriculture, autonomous driving and intelligent mining. China News Service highlighted an apple-orchard frost-prevention system and a 5G-A intelligent-mine project as examples of computing-linked deployment beyond generic large-model training. 25

From selling racks to selling tokens

The commercial goal is to move beyond leasing cabinets or reserving fixed servers. China Mobile said it is exploring token-based pricing for AI inference, model fine-tuning and agent hosting, allowing customers to pay according to consumption. 8

A token is a basic unit processed by an AI model. For a buyer, token-oriented services could be simpler than securing infrastructure, choosing accelerators and managing capacity directly. For the operator, the value proposition becomes the ability to dispatch workloads efficiently across hardware, power availability and network conditions.

That is why Qingyang’s “token factory” language is more than branding. The city’s stated direction is an integrated chain from green electricity to computing power to tokens. 11 Zhihui Blueprint has likewise presented its project as a move from selling cabinets toward selling tokens, with users drawing on pooled capacity and paying as they consume it. 36

What will determine whether the model works

Qingyang has assembled several ingredients of a compute-utility model: a rapidly expanding AI-focused cluster, lower reported power costs, renewable-energy integration, cross-region networking and an expanding supplier-and-customer base.

But the transition from racks to token services is still an operational challenge, not a guaranteed outcome. It requires consistently available power and networks, high utilization, interoperable scheduling across heterogeneous systems, strong data security and billing that customers can understand and trust. The clearest takeaway is that Qingyang is trying to make AI compute behave more like a utility: centrally produced, dynamically dispatched and bought on demand.