Research
2026-07-31
DeepSeek Plans $35B 1GW AI Data Center, Shifting from Rented to Owned Infrastructure
Summary
Key Takeaways
DeepSeek is advancing plans to build a hyperscale AI data center in Ulanqab, Inner Mongolia, with a planned capacity of 1GW, making it one of the largest AI data centers globally. The project is estimated to cost approximately $35 billion. The company has started hiring for senior engineering roles in project planning, construction management, electrical engineering, cooling systems, and commissioning, indicating the plans are beyond concept stage. Ulanqab offers low electricity prices and a cool climate to reduce cooling needs, aligning with China's Eastern Data Western Computing national strategy. Bernstein estimates that over 60% of the investment in a 1GW AI campus goes to GPUs, networking equipment, and storage. This project marks DeepSeek's shift from relying on third-party compute providers (previously reported to use about 20,000 chips rented from Alibaba to train the Kimi K3 model) to building dedicated infrastructure. This transition occurs amid escalating global chip export controls, highlighting Chinese AI companies' strategic adjustments in accessing advanced semiconductors. The open-source Kimi K3 model has previously matched top Western models in multiple benchmarks.
Why It Matters
DeepSeek's shift to self-built 1GW data center is a defensive move against US chip export controls, but it locks the company into domestic GPU ecosystems (e.g., Huawei Ascend), which still lag in linear scaling efficiency and interconnect bandwidth (e.g., HCCS vs NVLink), potentially causing longer training times and tail latency issues. The 1GW scale faces physical constraints in power quotas and cooling water; even with cool climate, large-scale liquid cooling requires significant water resources, risking environmental approval delays. DeepSeek has not disclosed its network fabric choice; if using InfiniBand, supply chain risks; if RoCEv2, it must tackle PFC/ECN congestion control. Enterprises relying on DeepSeek's model services should assess its infrastructure's elastic scalability and disaster recovery, as a single data center poses a single point of failure.
PRO Decision
[Vendors] Competitors like Alibaba Cloud and Huawei Cloud should capitalize on DeepSeek's massive capital expenditure and long construction timeline by promoting their elastic GPU cloud services with multi-tenant isolation and optimized networking (e.g., RoCEv2 with DPU acceleration) to reduce the need for customers to build their own infrastructure.
[Enterprises] Enterprise CIOs and architects should conduct zero-trust technology audits on DeepSeek's model services, demanding cross-cloud portability to avoid vendor lock-in. They should also assess the single point of failure risk of a single data center and require multi-region deployment commitments.
[Investors] Investors should see through the PR spin: DeepSeek's self-built data center is a defensive vertical integration against chip export controls, leading to soaring capital expenditure and depreciation pressure. The long-term gross margin of model services may suffer, and domestic GPU performance gaps could hinder model competitiveness. Monitor GPU supply chain diversification and network fabric choice as key risk factors.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)