Why Alibaba’s Ulanqab Data Center Is Emerging as a Hub for China’s AI Compute

Ulanqab illustrates a central shift in AI infrastructure: the best location for compute does not need to be inside the largest user market. Alibaba Cloud’s Inner Mongolia campus is close enough to serve Beijing with reported round-trip latency of roughly 4 milliseconds, yet it benefits from cheaper electricity and a climate that lowers cooling needs. That combination can make it a practical base for both AI inference and large-scale training. 2425

Distance matters less when the network is fast enough

Ulanqab is about 350 kilometers from Beijing, but Alibaba data-center operations chief Wang Zhaoyang told local media that data transmission between the two is about 4 milliseconds. He described that delay as acceptable for conversational AI applications. 25

That distinction is important. Physical distance is not the same thing as application delay: a well-connected regional campus can serve a nearby metropolitan market if its network transit is small relative to the rest of the response path. Ulanqab’s two direct fiber links to Beijing have reportedly reduced one-way latency to less than 2.1 milliseconds. 35

Power and cooling create the economic case

AI data centers consume enormous amounts of electricity, so the site’s power economics are foundational. Reports cite local electricity prices of roughly RMB 0.32–0.35 per kWh, with renewable energy accounting for about 90% of the local supply in one reported estimate. 2125

The climate adds a second advantage. Ulanqab’s annual average temperature has been reported at about 4°C, and local reporting says natural cooling can be used for much of the year. 2435 For an operator running dense accelerator clusters, less mechanical cooling can reduce operating costs alongside lower power prices.

This is also more than a general renewable-energy claim. A separate Ulanqab low-carbon computing-base project, integrating generation, grid, load and storage, entered operation in 2025. It is designed to supply green power directly to data-center loads and is projected to generate 848 million kWh of green electricity annually for self-use. 39

CUBE 5.0 makes deployment speed part of the strategy

Alibaba’s CUBE 5.0 architecture is intended to solve a different bottleneck: how quickly a data center can be made ready for IT equipment. Rather than install critical systems sequentially on site, the approach modularizes power supply, cooling, security, intelligent management and fire protection so components can be produced and prepared in parallel. 34

Alibaba says this can compress delivery of the underlying AI-data-center infrastructure to 100 days and lower construction cost by about 10%, compared with conventional delivery cycles of six to 12 months in China. 2534

The claim needs a precise reading: the 100-day measure runs from a basic plant or stacked containers to the point before servers and network equipment enter. It is not necessarily the full elapsed time from a project decision to a functioning AI service. 3234

CUBE 5.0 is also designed for changing hardware requirements. Alibaba says the architecture uses high-voltage DC power and supports both air and liquid cooling, with compatibility for heterogeneous chips and at least three generations of chip iteration. 34

The real constraint is not the building

Fast construction and inexpensive electricity cannot create leading accelerators. The availability of suitable chips, memory, interconnects and mature software remains a limiting factor for AI clusters.

Alibaba has begun offering Lingjun supernode instances based on its Zhenwu M890 chip in the Ulanqab region. 1 The M890 is presented as a combined training-and-inference accelerator, with Alibaba reporting 144 GB of memory and 800 GB/s of inter-chip bandwidth. 11

That helps expand domestic hardware options, but it does not settle the question of capacity quality. Building a competitive AI cluster depends on sustained component supply, networking, software tooling, reliability and the ability to run models efficiently across many accelerators—not simply on the number of chips delivered.

Why a GPU shortage can coexist with idle data centers

China’s AI-compute market has produced apparently conflicting signals: demand for high-quality capacity can be intense even while some data centers are underused. A 2026 industry report cited overall smart-computing-center utilization below 40%, while earlier reporting described significant underuse in parts of the market. 310

These figures should not be treated as a single, universal utilization rate. Measures differ by period, geography, chip type and whether they count installed, rented or actively used capacity. Still, they point to a structural issue: capacity can be idle if it is in the wrong place, uses chips unsuited to the workload, lacks sufficiently capable networking or software, or has no durable customer demand. 89

For Ulanqab, the commercial test is therefore utilization rather than announced capacity. Its location and construction model improve the cost and speed of supplying compute; they do not guarantee that every accelerator will be productively scheduled.

Alibaba’s “one cloud, many chips” challenge

Alibaba’s broader approach is to make multiple accelerator types available through a shared cloud platform rather than rely on a single hardware supplier. This makes the software and systems layer crucial: cluster architecture, networking, compilers, runtime software and model optimization have to work across different chips.

Alibaba says its T-Head (Pingtouge) Zhenwu series had cumulatively shipped more than 560,000 units to more than 400 customers across over 20 industries as of May 2026. 711 Those shipments demonstrate growing deployment, but they are not by themselves a measure of frontier-model performance or a substitute for a mature heterogeneous-compute stack.

What Ulanqab shows about AI infrastructure

Alibaba’s Ulanqab campus has a coherent physical logic: low-cost, renewable-heavy electricity; a cold, dry climate; network latency to Beijing measured in a few milliseconds; and an infrastructure design meant to shorten build cycles. 212534

Its longer-term significance depends on execution. The winning AI-compute hub will not merely be the place that builds the most capacity at the lowest cost. It will be the place that reliably matches compatible, well-networked chips with real workloads—and keeps those systems productively utilized.