Estimate the ENERGIZED AI-accelerator compute capacity of Chinese entities as of
2026-06-30, in gigawatts of IT load. I am reconciling a global AI-capacity model
in which US/Western-attributable capacity is independently built; the China line
is currently a placeholder of 9.0 GW. Your job is to confirm, revise, or refute
that number with evidence — not to agree with it.
## The quantity (frozen definition — do not answer a different question)
- BOTH training AND inference capacity are in scope: "accelerator compute" means
the chips regardless of workload. Training clusters are typically the largest
single consumers — excluding them is an instant failure. If evidence permits,
split training vs inference in the owner decomposition (graded); note that
export-compliant NVIDIA (H20-class) is inference-oriented silicon while
domestic training runs largely on Ascend — your split should be consistent
with your Method A chip mix.
- ENERGIZED only: racked, powered, running or available-to-run accelerators.
NOT announced projects, NOT completed-but-dark shells, NOT chips in inventory,
NOT pipeline/"planned by 2027." China has a documented history of idle/dark
datacenter capacity — treat "built" and "energized" as different states, and
idle-but-powered GPU halls as a separately flagged line.
- IT LOAD, not facility power. If a source gives facility MW, convert and state
the PUE you assumed. If a source gives chip counts, convert and state your
W-per-accelerator all-in figure (accelerator board power x host/network/cooling
overhead — state both factors).
- SCOPE: workload capacity owned/controlled by Chinese entities — hyperscalers
(ByteDance, Alibaba, Tencent, Baidu), AI labs (DeepSeek, Zhipu, Moonshot,
MiniMax, 01.AI, StepFun), telcos/state clusters (China Mobile/Telecom/Unicom,
regional "intelligent computing centers"), other enterprise. Report mainland
capacity and Chinese-owned OVERSEAS capacity (e.g., ByteDance renting or
building in Malaysia/Singapore/Middle East) as SEPARATE lines — I will decide
which to include.
- Attribute by WORKLOAD OWNER, not by host: ByteDance capacity hosted in a telco
facility counts once, under ByteDance. State your rule when ambiguous.
The 9.0 GW placeholder is refuted; the energized AI-accelerator compute capacity
of Chinese entities as of June 30, 2026, is 5.05 GW (Mainland China: 4.4 GW,
Chinese-owned Overseas: 0.65 GW).
White smoke