ByteDance Plans 6GW AI Data Center as Inner Mongolia Bet Tops $27.8B

A single infrastructure project in China's wind-swept north is set to concentrate more computing power than most nations possess, as ByteDance doubles down on a capacity race that is reshaping the global AI landscape.

ByteDance plans to build a 5-to-6 gigawatt AI data center campus in Ulanqab, Inner Mongolia, with delivery targeted for early 2028, according to a report by the South China Morning Post published September 3. The project, when combined with the company's existing Inner Mongolia footprint, would push ByteDance's total regional compute capacity to between 6 and 7 GW — and its cumulative capital commitment in the region past RMB 200 billion (US$27.8 billion).

The announcement underscores a strategic inflection point: Chinese technology giants are no longer treating AI infrastructure as an operating cost but as a sovereign-grade strategic asset, committing capital at a scale that rivals national grid investments.

Scale Redefines the Benchmark for AI Infrastructure

To contextualize the ambition: 1 GW of AI data center capacity can house approximately 500,000 H100-class GPUs by industry convention. At 6 GW, the new Ulanqab facility would theoretically accommodate up to 3 million GPU-equivalent units — a figure that places the project among the largest single AI compute concentrations on the planet.

For reference, a large nuclear power plant carries an installed generation capacity of 1 to 2 GW. ByteDance's new campus, in IT load terms alone, is the functional equivalent of three to six nuclear reactors running exclusively to power AI computation.

Actual electricity draw will be substantially higher. Factoring in cooling, power conversion losses, and facility overhead, a 6 GW IT load typically translates to more than 10 GW of total grid demand — a stress test for any regional power network.

Ulanqab's Structural Advantages Drive Site Selection

ByteDance's choice of Ulanqab is neither arbitrary nor purely political. The city sits at the intersection of three hard infrastructure advantages that directly compress operating costs.

Thermal environment: Average annual temperatures hover around 4°C, enabling natural air-side cooling for a significant portion of the year. Data centers in the region routinely achieve a Power Usage Effectiveness (PUE) ratio below 1.15 — meaningfully better than coastal Chinese hubs and well ahead of the global average of approximately 1.5.

Energy cost and green supply: Inner Mongolia hosts China's largest wind and solar generation base. Industrial electricity tariffs on the Mongolian Western Grid rank among the lowest nationally, while abundant renewable energy certificates allow hyperscalers to meet corporate carbon commitments — an increasingly non-negotiable requirement for companies with global ESG obligations.

Latency proximity to Beijing: Ulanqab sits roughly 300 kilometers from the capital, keeping fiber-optic round-trip latency within 5 milliseconds. The city already absorbs 75,000 petaflops of compute demand overflowing from Beijing data centers, making it the country's largest AI training cluster by some metrics.

As of the first half of 2026, Ulanqab had 89 operational data center projects with a running compute capacity of 165,000 petaflops (P), of which intelligent compute accounts for more than 90% of the mix. Tenants training large language models on that infrastructure include DeepSeek, Alibaba Group, and Baidu.

ByteDance's Inner Mongolia Position Already Runs Deep

The 5-to-6 GW project is not ByteDance's first capital deployment in the region. The company previously committed RMB 70 billion (US$9.7 billion) to a 1 GW AI Data Center (AIDC) cluster in Ulanqab — with a planned capacity of 200,000 P and partial facilities already in production — and a separate RMB 12 billion (US$1.67 billion) Volcano Engine intelligent compute center in Hohhot's Helinger district. Combined, these two prior commitments total RMB 82 billion (US$11.4 billion).

Corporate structure signals intent. In the first half of 2026, ByteDance registered four wholly-owned technology subsidiaries in Inner Mongolia with aggregate registered capital of RMB 4.5 billion (US$625 million). One entity — Ulanqab Yanbei Zhiwei Technology — carries RMB 400 million (US$55.6 million) in registered capital and appears purpose-built for the new campus.

Aggregated across all disclosed commitments, ByteDance's Inner Mongolia compute buildout now represents a potential RMB 200 billion-plus (US$27.8 billion) capital program — a figure that moves the project from corporate infrastructure into the category of regional industrial policy.

Compute Scarcity Drives the Arms Race Logic

ByteDance's capital aggression reflects the competitive arithmetic of frontier AI development. The company's Doubao large language model, multimodal systems, AI Agent frameworks, and Jimeng AI video generation tool are all consuming compute at accelerating rates. Video generation models, in particular, carry training compute requirements estimated at tens of times those of text-only models — and Jimeng's user base has expanded sharply in 2026.

The strategic calculus is straightforward: compute capacity determines model scale, which determines iteration speed, which determines market position. OpenAI, Microsoft, and Google are pursuing the same logic in North America and Europe. Domestically, Alibaba and Tencent are executing parallel buildouts. For ByteDance, the 2028 delivery window is not incidental — it aligns with the expected training cycle for the next generation of frontier models.

Industry Concentration Raises Infrastructure Risk Flags

ByteDance is not building in isolation. Publicly disclosed AI-related data center investment commitments in Inner Mongolia from domestic technology companies total approximately RMB 163.55 billion (US$22.7 billion). GDS Holdings, Huawei, Alibaba, and Baidu all have projects in the pipeline. Ulanqab alone has signed 67 data center projects with total committed investment of RMB 260 billion (US$36.1 billion) and a completed standard rack capacity exceeding 500,000 units.

That concentration creates three identifiable risk vectors.

Grid stress: Aggregated IT loads of this magnitude — potentially exceeding 10 GW of total power draw in a single city cluster — will test Inner Mongolia's transmission and distribution infrastructure, even in a province with structural power surplus.

Water scarcity: Ulanqab sits in a semi-arid zone with structurally constrained freshwater resources. While modern hyperscale designs increasingly rely on indirect evaporative or air-side cooling to minimize water consumption, the sheer density of planned construction introduces cumulative pressure that operators and regulators have not yet fully stress-tested.

Utilization risk: The fundamental bull case for this capital deployment rests on sustained AI demand growth through 2028 and beyond. If model efficiency improvements — smaller parameter counts achieving equivalent performance — accelerate faster than expected, or if enterprise AI adoption plateaus, the region faces the prospect of stranded assets at an unprecedented scale.

Ulanqab Emerges as China's AI Compute Capital

The data points converge on a single conclusion: Ulanqab is transitioning from a regional logistics and agricultural hub into the central node of China's AI training infrastructure. ByteDance's 5-to-6 GW commitment is the largest single announced project in that transformation, but it is one component of an investment wave that will fundamentally alter the city's economic identity over the next three years.

Whether the demand materializes to justify the supply is the defining question for the sector. For now, every major Chinese AI platform is placing the same bet — that compute scarcity, not model architecture, will determine the winners of the next phase of the AI competition.

Related Coverage:

ByteDance, Alibaba Pull AI Agents as Regulation Reshapes China’s AI Market