DeepSeek plans to build a 1GW large-scale AI data center: competition sinks from the model layer to the computing power base

According to market news, DeepSeek plans to build a large-scale AI data center in Inner Mongolia, with an additional computing power of approximately 1GW. At present, the project is still in the market news/media report level, and the official has not officially disclosed the details - but the strategic signal released is clear enough: the competition of large model manufacturers is sinking from the model layer to the computing power base.

Why Inner Mongolia

Inner Mongolia is one of the core locations for large-scale computing power centers in China. It relies on abundant green power resources and low electricity prices, and is also in line with the national layout of "digitizing in the east and computing in the west". For a company like DeepSeek that is based on "low cost + open source", green electricity + low electricity prices mean that the electricity costs of training and inference can be further internalized - this account directly supports its extremely cost-effective pricing strategy.

The flywheel from "selling models" to "building computing power"

The approximately 1GW of new computing power is at the hyperscale data center level and will significantly enhance DeepSeek’s own training/inference base. Self-built data centers not only bring about lower costs, but also greater autonomy over computing power supply, scheduling, and electricity prices, thus forming a self-reinforcing flywheel of "computing power-model-price": the cheaper the computing power, the lower the model pricing, and the more users, which in turn supports greater investment in computing power.

From an industry perspective, large model competition is no longer limited to parameters and capabilities—self-built computing power, dispatch efficiency, and electricity price costs are becoming the core variables of a new round of competition. Leading players are investing in upstream infrastructure. DeepSeek This step is to upgrade "cost leadership" from a strategy to an infrastructure-level moat.

Several directions worth tracking in the future:

1. Project Confirmation and Timetable: Whether the official has disclosed the project details and construction period.
2. Support for the V4 series: How 1GW of computing power can be transformed into practical advantages for training and inference.
3. East-to-West Computing and Green Power Implementation: Policy support and the pace of advancement of power supply.