HANGZHOU, China—The job posting appeared on DeepSeek’s careers page on a Tuesday with a title the company had never used before: IDC design and planning engineer. The description said the work would cover everything from site selection to construction drawings, and that successful candidates would take part in building infrastructure from the megawatt scale to the gigawatt scale.
The listing, posted June 9, is the clearest signal yet that DeepSeek, the Chinese AI lab that startled the industry with its low-cost models, is moving from renting compute to owning it. The company, which has relied on third-party data centers and cloud providers for most of its computing, is now hiring the engineers who design and build data centers, the people who lay out power distribution, cooling systems and server halls.
The timing is not coincidental. DeepSeek is in the middle of its first external financing round, with a valuation reported at about 350 billion yuan, roughly $48 billion. Potential investors include Tencent and Alibaba, according to people familiar with the matter. The hiring suggests how founder Liang Wenfeng intends to spend the money: not on marketing, and not primarily on headcount, but on physical computing infrastructure that the company will own outright.
The scale of the ambition is what sets it apart. A gigawatt of data-center capacity is roughly the electricity draw of a large city or a single nuclear power plant. Before the AI boom, the largest hyperscale campuses operated by Google, Microsoft and Amazon typically ranged from tens of megawatts to a few hundred. Only a handful of projects in the world, most of them in the United States, have been planned at the gigawatt scale.
DeepSeek’s shift has been building for months. The company has been recruiting operations and maintenance engineers for an intelligent computing center in Ulanqab, a city in Inner Mongolia chosen for its cheap land, cool climate and access to renewable power. The new design and planning role, based in Hangzhou, signals that the company is moving into the front-end engineering phase, the stage that typically precedes construction by eighteen months to three years.
The move reflects a strategic judgment about the AI industry’s trajectory. DeepSeek shocked the world in early 2025 with models trained at a fraction of the cost of rivals, and its open-weight releases forced Western labs to cut prices. But as model capabilities converge, the company’s leadership appears to believe that infrastructure is becoming the deciding factor. Renting compute puts a lab at the mercy of cloud providers and leaves it exposed to price changes and capacity shortages. Owning data centers, transformers and substations is the long-term answer.
China’s regulatory environment reinforces the logic. The government has pushed domestic AI companies to build self-reliant computing capacity, partly to reduce dependence on imported chips and partly to keep data and infrastructure within national control. DeepSeek’s self-build program dovetails with state policy, even as the company insists it operates independently.
The engineering requirements in the job description are telling. The posting emphasizes liquid cooling, high-density power distribution, digital-twin simulation and automated operations, the technologies of a next-generation AI factory rather than a traditional server room. DeepSeek appears to be designing for the density that GPU clusters require, where a single rack can draw more power than a suburban street.
The challenges are formidable. Building a gigawatt-scale data center in China means securing grid connections, which can take years, and lining up equipment supply chains for transformers, cooling systems and networking gear. Analysts note that power and land approvals, not construction, are usually the binding constraint, and that DeepSeek’s projects may not deliver capacity until 2028 or later.
Rivals are moving in the same direction. ByteDance has been expanding its own data-center operations, and other Chinese AI companies have begun buying land and power rights. The race to build infrastructure is becoming as competitive as the race to build models, and the winners may be the companies that secured land and electricity early.
The infrastructure push also reshapes how the industry reads DeepSeek’s economics. The lab’s famous efficiency, its ability to train competitive models on modest budgets, was built on rented hardware and clever algorithms. Owning gigawatt-scale campuses changes the cost structure: heavy capital spending up front, depreciation spread over decades, and a fixed base of power costs that falls only as utilization rises. Analysts who follow the company say the shift signals confidence that demand for DeepSeek’s models, at home and abroad, will grow into the capacity.
For DeepSeek, the investment is a bet that its models will keep growing in capability and that demand for its compute will justify the capital outlay. The company’s founder has said little publicly about the plans, but the job postings speak for themselves: DeepSeek is no longer content to rent the future. It is building it, one construction drawing at a time.
