China Mobile triples compute at its Hohhot hub to 20 100 petaflops

China Mobile, one of China's three largest state-owned telecom operators, has finished expanding its biggest intelligent computing center in Hohhot, the capital of Inner Mongolia. The site's total performance reached 20 100 petaflops, or about 20.1 exaflops. That is three times its previous capacity. Inner Mongolia Daily reported the figure on Monday.

What we know about the scale

Simple arithmetic: if the growth is exactly threefold, the hub delivered about 6.7 thousand petaflops before the expansion. Now a carrier-run facility, not a state laboratory, is talking about exaflop-scale compute. For a telecom, that is a notable shift: China Mobile looks less and less like a connectivity provider and more and more like a supplier of computing for AI.

Key figures from the report:

• Site: China Mobile's largest intelligent computing hub, Hohhot, northern Inner Mongolia.
• Final capacity: 20 100 petaflops (about 20.1 exaflops).
• Growth: three times the previous level.
• Status: the expansion was recently completed.
• Source: Inner Mongolia Daily, Monday.

The figure needs a caveat. Chinese "intelligent" computing centers are usually rated in petaflops at low precision, in formats such as FP16 or FP8, which are used for training and inference of models. These are not the numbers used to rank supercomputers on the TOP500 list, which relies on FP64. The 20 100 petaflops figure cannot be compared directly with Western supercomputers. The published excerpt of the report does not specify the calculation precision.

Why Inner Mongolia

The choice of region is no accident. It is one of the key sites in the state program "Eastern Data, Western Computing." The idea is to move compute-heavy workloads out of congested coastal megacities and into the western and northern provinces. Energy and land are cheaper there, and the cool climate lowers cooling costs. For training large models, which runs for weeks and does not need minimal latency to the user, that trade-off works.

Beijing also explicitly ties digital infrastructure to its AI ambitions. The Hohhot expansion fits that logic: operators build the capacity, and model developers, cloud providers and government agencies then rent resources on it.

What the report leaves out

The main questions remain unanswered. The available text does not say which accelerators the hub is built on, whether domestic chips or something else. It gives no investment figure, no utilization rate, no list of tenants and no power consumption. These are the parameters that show whether 20.1 exaflops will turn into compute that is actually used or remain a nice number in a report. In China, built centers have more than once sat idle for lack of customers or because of software incompatibility.

What to watch next

For anyone following the Chinese AI market, four things are worth checking:

• Whether China Mobile publishes public pricing or announces customers for the Hohhot hub. That will show whether the operator sells capacity to the market or mostly serves its own needs.
• Which accelerators are disclosed in future reports. The answer determines how resilient the project is to chip export restrictions.
• How the operator will measure capacity going forward, with or without a stated precision. Without it, comparisons with other hubs say little.
• Whether China Telecom and China Unicom announce similar expansions. If they do, this is a systemic strategy of the state operators, not a single project.

For now, the public data supports one confident conclusion: the site has become three times more powerful, and Beijing keeps investing in the physical foundation for AI.