The Global Race for Cheap Power: Where AI Data Centers Should Actually Go
pravchaw · reddit · 2026-09-12
A long post maps global electricity costs against AI data center siting.
The subsidy illusion: residential rates in Iran, Libya and Ethiopia run as low as $0.003–0.007/kWh, but stem from government subsidies or domestic hydro, and come with underinvestment and unreliable grids — a "subsidy trap."
Unsubsidized reality (LCOE): the ranking flips — Saudi Arabia under $0.011/kWh ($11/MWh), Chile consistently below $0.025/kWh (Atacama has the world's highest solar irradiance), and India/Australia among the lowest for new utility-scale builds.
Training vs. inference: power is 40%–60% of an AI data center's opex. Training runs for months, needs gigawatts of continuous power and doesn't care about millisecond latency, making remote cheap-energy hubs like the Chilean desert or the Middle East economically sensible. Inference needs ultra-low latency; serving New York from South America or the Gulf adds 100ms+ and ruins UX.
Real bottlenecks: AI chips can't sleep when the sun goes down, so 24/7 operation on solar requires heavy investment in battery storage, geothermal or backup natural gas — the main hurdle for hyperscalers in low-cost regions.
More from Infra
- Wafer launches 'most comprehensive' AI performance engineering repo, starting with Transformer inference deep-dive — ycombinator · 2026-09-12
- Together serves 23%-30% of all OpenRouter traffic for GLM 5.3 models — zhyncs42 · 2026-09-12
- VCs float 'hardware revenue derivative': fund compute costs via revenue share, not equity — ns123abc · 2026-09-12
- Relace hits 1T tokens/day on OpenRouter, serving 37% of DeepSeek v4 Flash traffic — stuffyokodraws · 2026-09-12
- Yutori's Navigator n2 runs browser agents at $1.46 per task on OSWorld 2.0 vs $13-$40+ for frontier models — DhruvBatra_ · 2026-09-12
- Auto-derived FlashAttention with SMEM and tensor core assignment shown off — vtabbott_ · 2026-09-12