China's AI Data Center Boom: 24 GW Capacity, 12-Month Builds, and Five Giants' Divergent Strategies

Market Scale: Global Number Two with a Massive Expectation Gap

According to SemiAnalysis's building-level, bottom-up survey tracking over 1,000 data centers across 60-plus operators, the United States leads with 56 GW of IT data center capacity as of end-2026, while China ranks second with more than 24 GW of live IT capacity—already exceeding EMEA and the rest of APAC combined. An additional ~20 GW of projects have firm timelines, and another ~30 GW have been announced but not yet broken ground, indicating enormous future growth potential.

Previous public estimates of China's AI compute capacity varied by up to 15×, often painting a picture of high vacancy. The report identifies the core contradiction: traditional retail colocation cabinets sit empty, yet high-power wholesale capacity purpose-built for AI workloads is in short supply.

Construction Model: Chinese Speed and Cost Advantage

The bottlenecks for data center construction differ fundamentally between the two countries. In the U.S., the primary constraint is power-grid interconnection approval, which can take 12–13 months. In China, the limiting factor is AI chip imports; land, power, labor, and permitting are abundant. As a result, a standard 100 MW AI-era data center can be delivered in 12 months, and approvals at East-Data-West-Computing hubs can be compressed to 3–6 months. Three engineering modes enable this velocity:

Pre-built civil works: Colocation providers bear the risk by completing shell construction ahead of cloud-vendor tenders, enabling T+6-month delivery.

Large-span steel structures in the west: Replacing multi-story concrete buildings in the east with low-rise, long-span steel frames allows foundation work and factory fabrication of steel components to proceed in parallel, drastically cutting on-site time.

Modular prefabrication: Tencent's T-Block container modules and Alibaba's CUBE 5.0 "100-day data center" shift mechanical and electrical work to off-site factories for parallel production, shortening commissioning cycles.

Demand Side: Five Hyperscalers with Divergent Compute Strategies

AI compute demand in China is highly concentrated among five companies—ByteDance, Alibaba, Tencent, Baidu, and Huawei—each pursuing a distinct approach:

ByteDance: Largest single procurements; lease-first foundation accelerating into self-build. Driven by Doubao (345 million MAU), the Seedance video model, and Douyin/TikTok global traffic. 2026 AI infrastructure budget raised above $30 billion. Export controls push large high-end GPU deployments to overseas clusters in Johor–Singapore–Batam, with rapid lease expansion abroad.

Alibaba: Closest to the AWS hyperscaler paradigm; hybrid construction mixing self-build, BOT, and third-party wholesale leases within the same campus to reduce capex pressure. Simultaneously develops Qwen foundation models and T-Head custom chips.

Tencent: Strategically restrained, playing catch-up. Leverages WeChat's massive traffic entry but slowed AI investment pace. In 2022–23 consolidated from small urban leased sites into self-owned large parks.

Baidu: Pivoting from model vendor to chip vendor. Lowest compute capex among the five; capital increasingly directed to Kunlunxin custom silicon. GPU cloud revenue growing rapidly, with margin improvement tied to proprietary chips rather than raw scale.

Huawei: Over 90% self-build; operates three super cloud hubs in Gui'an, Ulanqab–Horinger, and Wuhu. Plays dual role as both compute provider and chip supplier.

Overseas Expansion and Long-Term Outlook

Top Chinese cloud vendors' overseas leased capacity is projected to approach 4 GW by 2029, with Southeast Asia the fastest-growing region. Export controls make overseas sites a critical channel for accessing high-end GPUs. ByteDance relies mainly on leases abroad; Alibaba is expanding in Europe, the U.S., and Southeast Asia; Tencent Cloud leverages game publishing as a beachhead into the Middle East and Southeast Asia.

The report forecasts that the five hyperscalers' combined compute capacity will grow from end-2023 levels to roughly 21 GW by 2030—more than a threefold increase. Annual new capacity additions of 3–4 GW are expected in 2026–2027, with an additional 6 GW of reserved projects awaiting groundbreaking.