AI Compute Economics: CapEx per GW Is the Wrong Metric
BenBajarin · x · 2026-07-17
Core View: When evaluating AI compute (Gigawattonomics), looking solely at "capital expenditure per GW" is a misleading metric.
- Business Model Differences: Priced equally by GW, rack sales, compute leasing, and internal cloud services are three entirely distinct businesses.
- ROI: Model estimates show that under a 10% cost of capital requirement, commercial GPU leasing yields an ROI of just 0.9% over 6 years.
- Key to Competition: Revenue density does not equal economic return. Whether AI compute can recoup its capital costs depends on sound pricing and utilization. The true competitive advantage lies in whoever can sustain the highest "revenue per watt."
Related event: Gigawattonomics Sparks Debate on AI Infrastructure Payback(8 posts)→
More from Venture
- Moonshot AI reportedly targets a $50B round ahead of Hong Kong listing — ctjlewis · 2026-07-22
- Beth Kindig says her Q3 AI stocks report flags 5 winners up more than 100% YTD — Beth_Kindig · 2026-07-22
- AI’s economic forecasts are split by nearly a quadrillion dollars by 2035 — bittingthembits · 2026-07-22
- LLM Agent Trader pairs backtesting with model-generated trading signals — tom_doerr · 2026-07-22
- OpenRouter spend data suggests frontier models still capture most paying usage — thedealdirector · 2026-07-21
- AI is collapsing startup and operating costs as U.S. business formation hits a record — sanjaykalra · 2026-07-21