As of mid-2026, compiled public data puts Xinjiang's AI computing power at close to 30,000 P. Karamay, Hami, Urumqi and Changji are all accelerating construction of computing clusters, while Tacheng, Altay and Turpan are moving ahead with projects of their own.

Cheap green power, a climate well suited to running data centers, and steadily rising demand for artificial intelligence. On those terms alone, Xinjiang does have real advantages in building out a computing industry.

But one question matters more than how many halls have been built or how many petaflops have been racked up: once the facility is finished, who pays for it?

Cabinets installed, equipment powered on, a project formally in service — none of that means the business works. For the local governments and investors building out computing capacity, the number that actually needs to add up is whether the spending converts into a sustained stream of customer orders, and ultimately into a reasonable return on investment.

Scale Is One Story, Utilization Is Another

Start with Karamay. Public figures put the city's AI computing power at more than 17,000 P, spread across six large data centers and roughly 65,000 standard cabinets, with a utilization rate of about 72%. The workloads span cloud services, telecoms, oil and gas data, off-site disaster recovery and film and television rendering, with some capacity sold to customers in eastern China and overseas.

What stands out here is not just the scale. More importantly, Karamay has already built up a base of real customers and use cases: a Huawei Cloud data center, a China Mobile AI data center, a PetroChina data center, plus rendering and disaster recovery work. Together they form the commercial foundation of the local computing industry.

That shows Xinjiang's computing capacity is not without a market. The real question is which projects have already found customers willing to keep paying, and which still need to prove there is demand. The gap between the two could directly shape how these projects perform.

A 72% utilization rate is not a 72% profit margin. Whether a cabinet generates revenue, what price the contract carries, how steep equipment depreciation and operating costs run — all of that has to be calculated separately. But one point deserves attention: computing capacity can be expanded with capital, while a steady book of customer orders does not come automatically with the building. That is why judging a computing project means looking past the investment figure and the construction timeline to how much business it has actually won.

Power at 0.3 Yuan per kWh Solves Only Half the Cost Problem

Electricity is the advantage Xinjiang cannot avoid talking about. Industry estimates put the lowest direct-connection price for green power in the region at 0.3 yuan per kWh. Abundant wind and solar resources, combined with operating conditions well suited to data centers in parts of the region, hand the computing industry a cost edge.

But cheap power only means part of the cost base is favorable. It is not proof that a project will turn a profit.

Run the simple math. On the estimates used here, a 1,000 P AI data center running at full load consumes roughly 9 million kWh a year. If the actual procurement price falls by 0.1 yuan per kWh, with consumption unchanged, that translates in theory to 900,000 yuan in annual electricity savings.

That is a meaningful cost gap, but 900,000 yuan is not profit, and it is certainly not a return on investment. Beyond the power bill, a computing center carries construction costs, servers and related equipment, network bandwidth, day-to-day operations and equipment depreciation. If the cabinets have no customers, no electricity price is low enough to generate revenue on its own. Flip it around: with stable customers and high equipment utilization, a project can rest on firmer footing even if the power price advantage is smaller than expected.

So judging whether a computing project is competitive means weighing at least three things together: the electricity and other operating costs the facility will actually carry; how much customers will pay for computing services and how long those orders will last; and what is left over after construction spending, equipment depreciation, bandwidth and operations are deducted.

Cheap power solves the cost side. Orders solve the revenue side. Only when both add up does an advantage become profit.

Not Every Computing Workload Belongs in Xinjiang

Some workloads demand extremely fast response times. High-frequency financial trading, real-time industrial control, and cloud gaming and interactive services aimed at users in eastern China are all sensitive to network latency. The farther a data center sits from its users, the more carefully data transmission has to be engineered. Xinjiang's geography imposes a constraint here, and cheap power alone is no reason to assume the region can serve every kind of demand.

Another category looks entirely different. Large model training, film and television rendering, cold data storage, off-site disaster recovery and some scientific computing workloads are relatively tolerant of latency and care more about compute costs, operational stability and data security. These are the workloads worth going after.

AI training and schedulable compute workloads. Training large models typically requires substantial computing resources and can run for extended periods. If a job can be queued, paused or rescheduled, there is room to align the compute load with wind and solar output — running more workloads when renewable generation is high and supply allows, and adjusting schedules when power is tight. That is the commercial value of coordinating compute with electricity. Not every training job can be interrupted at will, however, and not every cheap kilowatt-hour can be used without conditions. Power supply, workload scheduling and customer contracts have to line up before the approach becomes workable.

Cold data storage and off-site disaster recovery. Data backup does not necessarily chase ultra-low latency, but it puts a premium on security, stability and long-term availability. These workloads have a shot at becoming the base business for a computing park, though customers will also scrutinize recovery capability, network connectivity, operational support, compliance requirements and service contracts. Meeting those conditions determines whether the work turns into long-term revenue.

Film and television rendering and content production. Rendering is clearly an offline workload: assets go up, the job runs, results come back, and users generally do not need real-time feedback along the way. Karamay has deployed roughly 20,000 rendering nodes and has contributed to the rendering of more than 3,000 film and television productions. That points to an existing business base in the field. The next step is not simply replicating data center halls, but continuing to win actual orders in film and television, animation, advertising and digital content production.

Digital demand from Xinjiang's own industries. Oil and gas exploration, coal chemicals and cotton-field remote sensing could also generate sustained demand for computing power. The advantage here is that the demand comes from local industry rather than depending entirely on customers elsewhere. If a computing center can solve real data processing and computation problems in production, it has a chance to build more targeted service capabilities.

Computing services aimed at Central Asia. Xinjiang's location creates room to offer digital services to Central Asia, but geography is not the same as orders in hand. Market size, customers' ability to pay, network connectivity, cross-border data rules and actual commercial demand all have to be verified one by one. It is a track worth cultivating early with customers and partners, but it should not be treated as current investment returns in the absence of actual orders.

These five categories share one trait: none depends entirely on instant response for users. But meeting that condition does not guarantee a project succeeds. Distance tolerance is only the first filter. Someone willing to keep paying is the second.

The Day the Lights Come On Is When Illusions Set In

The moment a computing project is most likely to create a false impression may be precisely when the facility is finished. The lights are on, the equipment is installed, the marketing material exists, and the computing capacity can be tallied up. From a business standpoint, none of that substitutes for revenue.

Picture two computing centers with similar conditions. Project A signed service contracts with customers in advance and can point to a defined usage plan and revenue source for part of its equipment. Project B spends the capital first and goes looking for customers afterward. Even if B has the larger facility and the cheaper power, it faces greater operational uncertainty: the construction capital is already committed, the equipment is already depreciating, operating expenses keep coming, and if customers fail to materialize, the investment struggles to convert into revenue on schedule.

That is why a computing project cannot be judged on construction progress alone. Contract quality matters too. Before investing, there are at least four questions worth asking.

  • Who are the customers? Are they confirmed, paying customers, or prospects with nothing more than stated interest? Signing, piloting and actually paying are different stages.
  • How long are the orders? Short-term trials, or contracts that can support long-term operations? Do the contract terms match the useful life of the equipment?
  • How much of the equipment will be used? What is the projected utilization rate based on? Another park's utilization rate cannot simply be adopted as your own target.
  • Can revenue cover costs? Once electricity, bandwidth, operations, depreciation and financing costs are factored in, is there any profit left?

Without clear answers to these questions, the bigger the investment, the greater the uncertainty that comes with it.

One more thing gets overlooked: outbound bandwidth from Xinjiang and network costs. A computing center needs more than power. It needs to bring data in and send results back out. For data-heavy workloads, bandwidth conditions and network costs shape both customer experience and operating returns. So the comparison cannot be limited to power prices, or to cabinet counts. The comparison that matters puts electricity, networks, equipment, customers and operating capability on the same balance sheet.

A 1,000 P Center Needs Three Sets of Books

Back to that 1,000 P AI data center. On the estimates used here, full-load operation consumes about 9 million kWh a year; a 0.1 yuan per kWh reduction in the actual power price implies a theoretical saving of roughly 900,000 yuan, assuming full load and identical consumption. But a complete project model needs at least three layers.

The revenue book covers contract values, the volume of computing services actually delivered, pricing structure, contract terms and whether the revenue is sustainable. The operating book covers actual power consumption, network bandwidth, equipment maintenance, staffing, cooling systems and other running costs. The investment book has to account for construction spending, servers and related equipment, depreciation, financing costs and future upgrade spending.

Only with all three sets of books in view can anyone reasonably assess the break-even point, the payback period and the capacity to absorb risk.

It is worth stressing that Karamay's roughly 72% utilization rate cannot be applied as a forecast for every computing project in Xinjiang. Customer mix, workload types, equipment configuration and operational standards differ from park to park, and the results can vary widely. The same goes for estimates of cheap power, electricity savings and regional cost differences, all of which need to be rechecked against each specific project.

A computing project cannot prove it makes money with a tidy set of projections. It has to prove itself with operating data that holds up under scrutiny.

From a Construction Mindset to an Operating Mindset

Xinjiang has green power, favorable climate conditions and local industries such as oil, gas and manufacturing that generate real use cases. Those are genuine foundations. But future competition will not hand a region an advantage simply because it has more cabinets.

For Karamay, the priority is deepening its oil and gas data, rendering, disaster recovery and training businesses while improving resource efficiency. For Hami and Yiwu, coordinating compute with electricity and building green computing capacity can be important directions, but expansion still has to match real orders. For Urumqi, Changji and similar cities, digitalizing local industry and serving computing demand within Xinjiang can provide meaningful support. For other regions still laying out projects, identifying clear customers and use cases before deciding on scale may be safer than chasing the largest possible build.

All of this points in one direction: shifting from a construction mindset to an operating mindset. Xinjiang does not need to win every computing workload, nor does it need to copy another region's development path. What is worth doing is identifying the markets where it has a cost advantage, workload fit and an industrial base, then validating investment with actual orders.

A computing center is not an asset that earns money the moment it is built. It needs technology and power, but also customers, contracts and operating capability. The next phase for Xinjiang's computing industry should move from "how much we built" to "how much is in use, who is buying it, and how much it earns."

A finished facility is where the project starts. Customers who keep paying are where the business starts. What ultimately separates the winners is not who has more cabinets, but who can keep more cabinets creating value.