DeepSeek Plans 1-Gigawatt AI Data Center in Inner Mongolia: What We Know About the Timeline, Location, and Chip Question

Capacity and Scale

The planned 1-gigawatt facility would be among the largest AI data center projects globally, comparable to the biggest hyperscaler builds by companies like Microsoft and Google. DeepSeek plans to build its own facility while also leasing additional capacity from other operators. A third-party compilation of local government approvals shows more than ten companies have already filed projects in Ulanqab, suggesting the area is becoming a concentrated compute hub in China.

Timeline

DeepSeek is pushing to bring at least part of the capacity online by the end of 2027 or early 2028. This is an aggressive timeline given the scale of construction and permitting required for a gigawatt-class facility.

The Critical Unknown: Chips

The single biggest unresolved question is what chips will power the facility. The U.S. has imposed export controls barring advanced Nvidia GPUs (e.g., H100, H200, Blackwell B200) from entering China. Some reports allege DeepSeek has previously obtained banned Nvidia Blackwell chips through intermediary channels in violation of U.S. sanctions, though this has not been confirmed by the company or U.S. authorities. It is unclear whether DeepSeek plans to rely on domestic alternatives (e.g., Huawei Ascend, Cambricon), stockpiled Nvidia chips, or gray-market procurement for this facility. The choice will determine whether the project is viable under existing export controls or represents a direct challenge to them.

Why Ulanqab? Strategic Location Advantages

1. Alignment with 'East Data, West Computing' Policy

Ulanqab is explicitly designated as a key hub under China's national 'East Data, West Computing' (东数西算) policy, which routes compute-intensive workloads from the coastal east to the resource-rich west. The policy provides favorable land prices, electricity tariffs, streamlined approvals, and grid interconnection priority for large data centers in approved western zones.

2. Cheap, Abundant Renewable Energy

Inner Mongolia is China's largest renewable energy base, with massive wind and solar farms. This helps DeepSeek secure low-cost green electricity — critical for a 1-GW facility whose largest operating expense will be power.

3. Cool Climate Reduces Cooling Costs

The region's cold, dry climate for much of the year allows for natural air or evaporative cooling, significantly lowering the facility's power usage effectiveness (PUE) compared to data centers in warmer coastal cities. Ulanqab's average annual temperature is around 4°C.

4. Land Availability and Proximity to Beijing

Ulanqab offers abundant, cheap land for large-scale campus builds while being close enough (350 km, roughly 3 hours by high-speed rail) to Beijing's tech talent and R&D centers.

What This Signals for DeepSeek

DeepSeek built its reputation on efficiency breakthroughs — achieving frontier-level model performance with far fewer GPUs than Silicon Valley competitors, training reportedly for under $6 million. This project signals a strategic shift:

• From capital efficiency to scale: The company is signaling that to compete at the frontier, it can no longer rely solely on algorithmic cleverness; it needs hyperscale compute capacity of its own.
• Preparing for next-generation models: A 1-GW facility suggests DeepSeek is planning to train and serve models an order of magnitude larger than its current offerings, likely requiring sustained clusters of 100,000+ accelerators.
• Infrastructure as competitive moat: DeepSeek appears to be moving from a 'lean startup' model toward a vertically integrated infrastructure strategy akin to OpenAI/Microsoft and Google/DeepMind, where owning compute is seen as essential to long-term competitiveness.

Bottom line: This is a strategic escalation. The chip question — which hardware will fill 1 GW — remains the most consequential uncertainty and will determine whether the project is viable under U.S. export controls or represents a direct challenge to them.