DeepSeek Is Hiring the Engineers Needed to Build Its AI Infrastructure

DeepSeek’s latest IDC recruitment points to a strategic change in how AI companies may secure compute. Rather than treating data centers as a commodity rented from third parties, the company is building expertise across infrastructure planning, construction, testing, and operations in Beijing, Hangzhou, and Ulanqab. 12

That does not prove DeepSeek will own every facility it uses. It does show that the physical systems behind AI—power, cooling, buildings, networks, and operational reliability—are becoming important enough to manage as part of the product stack.

The job categories reveal the strategy

DeepSeek’s listings seek specialists in electrical engineering, HVAC, automation, energy, communications, computer science, environmental engineering, and civil engineering. The roles cover infrastructure operations, planning and design, and on-site implementation. 34

This is a materially broader profile than a team focused only on buying cloud capacity. It covers the systems that determine whether an AI cluster can run at high density and remain reliable:

• Electrical and energy engineering: power delivery, substations, distribution, and energy coordination
• HVAC and environmental engineering: air cooling, liquid cooling, heat rejection, and environmental controls
• Civil engineering: sites, buildings, structural works, and construction delivery
• Automation and communications: monitoring, control systems, network connectivity, and operations
• Computing and infrastructure operations: integrating physical capacity with clusters, scheduling, and workloads

DeepSeek’s own recruitment language reportedly spans conventional data centers, AI computing centers, air- and liquid-cooling systems, and projects from megawatt to gigawatt scale. 78 The significance is less the scale implied by a job description than the breadth of capability the company is trying to assemble.

From rented compute to operational control

Reports describe the move as a departure from DeepSeek’s earlier asset-light reliance on third-party computing capacity. 12 Building internal infrastructure expertise can give an AI developer more control over cluster design, equipment compatibility, cooling architecture, workload scheduling, and expansion timing.

That control matters because AI performance depends on more than the number of chips installed. A cluster can deliver useful compute only when its power systems, cooling, networking, storage, software, and maintenance processes work together. DeepSeek’s recruitment description explicitly links electricity, cooling, networks, and computing systems to final compute output. 13

The strategic lesson is therefore narrower—and more defensible—than “every AI company must build its own data centers.” The real advantage is assured access and operational control. That can come through ownership, long-term leases, colocations, joint ventures, or tightly integrated infrastructure partnerships.

Why the locations suggest a split compute model

DeepSeek is recruiting for its IDC team in Beijing, Hangzhou, and Ulanqab. 13 Those locations are consistent with a distributed architecture rather than a single national campus.

Inland hubs such as Ulanqab can be attractive for workloads that tolerate network delay, including model training, large batch processing, and other asynchronous jobs. Reporting on Inner Mongolia’s computing buildout highlights the region’s role in green computing, while earlier coverage described Ulanqab data centers serving Beijing with network latency below five milliseconds. 2131

Metropolitan locations remain better suited to latency-sensitive inference and enterprise services, where proximity to users, networks, and data can matter more. This suggests a two-tier topology:

1. Inland, power-oriented clusters for training, fine-tuning, synthetic-data generation, and batch inference.
2. Urban or edge-adjacent capacity for interactive consumer and enterprise inference.

DeepSeek’s hiring footprint fits that hypothesis, but a recruitment notice cannot establish how the company will ultimately allocate workloads. 13

Ulanqab is becoming an AI-compute cluster

Ulanqab’s importance is not based on DeepSeek alone. Multiple reports describe a fast-growing concentration of data-center and intelligent-computing projects in the city and the wider Inner Mongolia region.

One report cited 84 signed data-center projects in Ulanqab by the end of 2025, including 81 intelligent-computing centers, with planned investment exceeding RMB 500 billion. 27 Other reporting has described nearly 100 projects and approximately 12.5 gigawatts of announced capacity, although the figures and definitions vary by source. 2324

The region’s appeal is the prospect of combining land, power, fiber connectivity, equipment suppliers, and specialist operations talent in one expanding ecosystem. Envision’s Galaxy Campus in Ulanqab was reported to have entered operation in August 2026, with a 120,000-square-meter facility and a claimed ability to run one million AI chips in parallel. Those capacity claims come from company and media reporting and should be treated accordingly. 202225

A cluster can reduce the friction of future deployments because infrastructure is shared across projects. But announced capacity is not the same as energized, installed, or fully utilized capacity. That distinction is essential when comparing the scale of China’s AI buildout.

What the ByteDance comparison does—and does not—show

The broader infrastructure trend also appears in reporting about ByteDance-related entities in Zhongwei, Ningxia. One July report said a ByteDance-owned company, Zhongwei Saishamingsha Technology, had been established with RMB 2.2 billion in registered capital and business activities including information-technology consulting, software development, systems integration, and sales of computer equipment. 52

That evidence does not verify the more specific claim that ByteDance incorporated two Zhongwei subsidiaries with registered capital of RMB 220 million and RMB 240 million, each with non-residential real-estate leasing scopes. Nor does it establish that the reported entity was created specifically to build a data center.

The safer conclusion is that major technology companies are exploring tighter control over land, buildings, equipment, and compute capacity. Separate reporting has described Chinese internet companies increasing strategic land purchases tied to core operations, but that broader trend should not be treated as proof of a particular ByteDance project. 58

The bottleneck is moving beyond the building shell

For conventional data centers, the building is a visible milestone. For dense AI clusters, the harder constraints may be the systems inside and around it:

• firm grid access and generation capacity;
• transformers, substations, and power distribution;
• liquid-cooling and heat-rejection systems;
• network links and cluster interconnects;
• water, environmental, and site requirements; and
• engineers who can commission and operate the systems as one platform.

Reporting on AI data-center development identifies electricity supply, high power density, and liquid cooling as increasingly important variables. 7 Inner Mongolia’s 2026 green-computing conference also showed how computing projects are being linked with equipment manufacturing, energy systems, and industrial investment: 12 announced projects carried a planned investment of RMB 186.46 billion. 2126

This helps explain why civil, electrical, HVAC, environmental, automation, and energy engineers are appearing in AI-company hiring plans alongside model researchers and software developers.

AI compute is becoming infrastructure—but not a universal ownership requirement

DeepSeek’s recruitment is best read as an infrastructure signal, not as final proof of a completed campus. The company appears to be building the organizational capability to plan, deliver, and operate AI data-center systems across multiple locations. 1416

That capability could strengthen its position when compute is scarce, clusters are difficult to configure, or high-density cooling and power access determine usable output. It also reflects a wider reorganization of AI economics: competitive advantage increasingly depends on securing the physical conditions in which models can be trained and served.

The comparison with railways, the power grid, and the internet is useful as a description of capital intensity and network dependence. But the durable moat is not necessarily owning a gigawatt campus outright. It is having reliable, expandable access to power and compute—and the operational control to turn that access into consistent AI service.