DeepSeek, the Hangzhou-based AI lab that rattled global markets with its R1 model in January 2025, is quietly building something new in China’s north. Job postings for data center engineers in Inner Mongolia have surfaced in recent weeks, pointing to a strategic shift: DeepSeek appears to be moving from reliance on third-party cloud providers toward owning and operating its own compute infrastructure. The move, if confirmed at scale, would mark a significant maturation in the lab’s operational model and raise important questions about its chip procurement strategy ahead of an anticipated V4 model launch.
Why Inner Mongolia? The Economics of Cold Air and Cheap Power
Inner Mongolia is not a random choice. The region sits at the intersection of two critical cost factors for large-scale AI compute: abundant, inexpensive electricity and a cold climate that dramatically reduces cooling costs. The region draws on a mix of coal-fired and wind power, making it one of the cheapest places in China to run power-hungry GPU clusters. Several of China’s largest data center operators, including China Telecom and GDS Holdings, have already established major facilities there for the same reasons. According to reporting by the South China Morning Post, the job postings signal a deliberate shift in DeepSeek’s infrastructure strategy.
For an AI lab that has built its competitive advantage on radical cost efficiency, DeepSeek’s R1 reportedly costs a fraction of comparable US models to train, owning the power and cooling infrastructure rather than paying cloud markup rates is a logical next step. The move aligns with the company’s broader strategy of optimizing every layer of the AI stack, from algorithmic innovations to hardware utilization. At scale, the cost difference between owned and rented compute can be enormous: hyperscalers routinely achieve 30–50% lower effective compute costs compared to cloud tenants.
From Cloud Tenant to Infrastructure Owner
The move signals a maturation in DeepSeek’s operational model. Until now, the lab has been understood to rely heavily on cloud compute from providers including Alibaba Cloud and Tencent Cloud, supplemented by access to Huawei Ascend clusters. Building proprietary data centers changes the calculus significantly: it locks in long-term compute costs, gives DeepSeek direct control over hardware procurement, and reduces exposure to any future cloud pricing shifts.
It also raises questions about DeepSeek’s chip strategy. The company has been notably opaque about its hardware stack. While it is known to use Nvidia H800 GPUs — the export-controlled version available in China before tighter restrictions — it has also been linked to Huawei’s Ascend 910B and 910C chips. Building its own data centers means DeepSeek will have to make explicit decisions about which chips to deploy at scale, a choice that carries both technical and geopolitical dimensions. Huawei’s Ascend chips are improving rapidly but still trail Nvidia’s H100 in raw performance; the tradeoffs between cost, availability, and performance will shape DeepSeek’s next-generation training runs.
V4 Delay and the Compute Question
The hiring push comes as speculation mounts about the delayed launch of DeepSeek V4. The model was widely expected in Q1 2026 but has not materialized, and industry observers have pointed to compute constraints as a plausible explanation. Training a next-generation frontier model requires sustained access to tens of thousands of high-end accelerators, a resource that is increasingly difficult to secure in China given US export controls.
If DeepSeek is building its own data centers, the timeline for V4 may be tied to when that infrastructure comes online. Inner Mongolia data centers typically take 12 to 18 months to build from groundbreaking to operational. Job postings for engineers suggest the project is in early-to-mid development, not imminent. This timeline suggests that DeepSeek is preparing for the long haul, prioritizing sustainable infrastructure over rushed model releases. It also suggests that V4, when it arrives, may be trained on substantially more compute than previous DeepSeek models, potentially enabling a step-change in capability rather than an incremental improvement.
What It Means for China’s AI Infrastructure Race
DeepSeek’s move is part of a broader pattern. ByteDance, Alibaba, and Baidu have all been aggressively expanding their own AI compute infrastructure in 2025 and 2026, driven by the same logic: as AI training and inference costs become a primary competitive variable, owning the underlying infrastructure is a strategic necessity, not a luxury.
What makes DeepSeek’s case distinctive is its size. It is a relatively small lab by the standards of China’s tech giants, and the decision to invest in proprietary data centers suggests its leadership believes it is building something large enough to justify the capital expenditure. That is a meaningful signal about the lab’s ambitions — and about how seriously it is taking the long game in the race to build China’s most capable AI systems. For the broader Chinese AI ecosystem, DeepSeek’s infrastructure push also reinforces a message that has been consistent since R1’s launch: the lab is not a one-hit wonder but a serious long-term competitor with the resources and strategic clarity to challenge both domestic rivals and global frontier labs.