AI research wiki
hardware · 2026-09-25 · Tier 1

SemiAnalysis: The Chinese AI Infrastructure Boom (China Datacenter Model)

SemiAnalysis: The Chinese AI Infrastructure Boom (China Datacenter Model)

Source: SemiAnalysis, 2026-09-25 (post), free preview of a paid model Raw: raw/rss/2026-09-25-semianalysis-the-chinese-ai-infrastructure-boom-introducing-the-semi.md

TL;DR

SemiAnalysis extends its building-level datacenter model across China's border, tracking 1,000+ facilities across 60+ operators. The market had settled on two assumptions, "China is big" and "China is empty," with published capacity estimates differing by 15x. The model finds over 24GW of Chinese datacenter capacity, larger than EMEA (14GW) and APAC ex-China (15GW), against the US at 56GW by end of 2026. That excludes ~20GW of dated pipeline and ~30GW announced. High vacancy is real, but it is a two-speed market: legacy retail racks sit near 60% utilization while wholesale AI buildings refill above 70%. BAT (Alibaba, Tencent, Baidu) capex hit $20B in 2Q26, more than double YoY, and all three posted negative free cash flow for the first time. ByteDance, which files no public accounts, occupies about a fifth of delivered capacity and rents nearly all of it.

flowchart LR
  E1[Era 1<br/>carrier hosting] --> E2[Retail colocation<br/>overbuild, vacancy]
  E2 --> AI{AI demand<br/>from 2023}
  AI -->|wholesale| W[AI halls<br/>back above 70%]
  AI -->|legacy| R[Retail racks<br/>near 60%]
  POL[Eastern Data,<br/>Western Compute] --> W
  W --> CAP[24GW+ delivered<br/>+20GW pipeline<br/>+30GW announced]
  EXP[Export controls<br/>on chips] -.->|constrains| W
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class E1,E2 input
  class AI decision
  class W,CAP output
  class R,EXP warn
  class POL aux

Key claims

  • Scale. China 24GW+ vs US 56GW, APAC ex-China ~15GW, EMEA ~14GW, LatAm ~2GW (2026YE).
  • Speed. China routinely delivers 100MW facilities in under 12 months, largely free of power constraints, labor shortages and protests. Modular builds have been standard since Tencent's third-gen design in 2014.
  • Who builds. State carriers still own a third of national capacity. GDS and VNET, the only US-listed Chinese landlords, signed 1.3GW of wholesale in 1H26 but captured barely a third of ByteDance and Alibaba orders from 2024 to 2026YTD.
  • Grid capex. The 14th Five-Year Plan ended 24% over blueprint. The 15th (2026 to 2030) adds another 40%, to over ¥5T ($746B).
  • Offshore. Chinese hyperscalers' overseas leasing doubles from 2026 to 2029 toward ~4GW, not counting hundreds of thousands of GPUs rented from Western clouds.

How this relates to prior wiki pages

  • Complements ClusterMAX 3.0 (09-24), which opened with "GPU supply has gone to zero" for Western GPU clouds. China's constraint is the mirror image: power and buildings are abundant, chips are the bottleneck under export controls.
  • Feeds compute-economics. Negative free cash flow at all three of BAT is the Chinese version of the hyperscaler capex story the wiki has tracked on the US side.
  • Context for the open-weights cadence. GLM 5.3 and Kimi K3, named in the opening, are the same models LLM Compressor v0.14 and Red Hat's NVFP4 checkpoints target this week. The infrastructure behind China's open-weight releases is larger than Western narratives assumed.

Gaps

This is a paid-model preview. Tenant-by-tenant detail, AI-specific capacity and chip counts are behind the paywall. Capacity in GW is IT power, which says nothing about how much of it holds current-generation accelerators.