AI data centers' power hunger fuels SMR startups; Weilan Zhidian bets on water-cooled small reactors
快鲤鱼 · wechat · 2026-09-03
AI data centers spark an SMR boom
Electricity can account for 50-60% of AI data center (AIDC) operating costs, and small modular reactors (SMRs) — with 90%+ capacity factors and the ability to connect directly to data centers, bypassing the grid — are becoming a hot power solution. Weilan Zhidian, founded by veteran Chinese nuclear engineers (CTO Xian Chunyu was chief designer of Hualong One), raised a Pre-A round from Northern Light VC and is developing its first reactor, WL-100.
Key points from the interview
- China vs. US power models: The US grid is fragmented into regional systems, pushing companies toward self-built power (Stargate's 15GW plan has stalled partly over electricity); China's unified grid makes supply chains longer and slower, though direct-connect models are now being explored.
- Domestic status: Demonstration projects like Shidaowan's HTGR are grid-connected and Linglong One may connect by year-end; new August regulations removed off-site emergency requirements for small reactors, and Beijing recently policies encourage nuclear direct-connect for AI data centers.
- Economics: SMR power costs 0.4+ RMB/kWh vs. 0.3 for large reactors, so the play is integrated energy savings — avoiding grid fees and using waste heat — not price competition.
- Technology: A deliberate choice of the most mature water-cooled route over HTGR/molten-salt designs to reach commercial operation fastest.
- Go-to-market: Coastal Chinese data centers and industrial parks, plus Belt-and-Road and Southeast Asian markets; the real signal will be the first approved data-center-plus-nuclear binding project.
More from Infra
- Magnitude open-sources a local-model inference server that plugs into Claude Code, Codex, Cline and more — magnitudedev · 2026-09-03
- Qwen3.6-35B at Q2_XXS: A GPU-Poor Laptop Builds a Zelda-like RPG in 24 Minutes — ML-Future · 2026-09-03
- Three weeks of llama.cpp optimizations: collecting best t/s for Qwen3.8-27B — pmttyji · 2026-09-03
- Running Physical AI Ops Means Mastering Everything from Roofing to Red Bull Logistics — MatthewChang · 2026-09-03
- NVIDIA's official speculative decoding guide is dominated by Chinese research teams — 新智元 · 2026-09-03
- Vector Ingestion at 50M Rows: The Pitfalls Your 1,000-Doc Prototype Won't Reveal — victorialslocum · 2026-09-03