AI overbuild math: 2027 compute wave may need $2.5T+ revenue, Jevon's paradox debate revisited
bookwormengr · x · 2026-11-01
The author revisits a viral calculation based on SemiAnalysis power-addition data: even pricing half the compute at DeepSeek V4.1 Flash inference rates (economics validated by the vLLM and SGLang teams), next year's online compute would require $2.5T in revenue to be consumed.
Key data from the charts:
- Global datacenter net capacity additions: 5.8GW (2023) → 9.8GW (2024) → 15.6GW (2025) → 25.0GW (2026E) → 51.2GW (2027E)
- If 50% of 2027 capacity (45.9GW) runs inference at $100B/GW/year, the "fair revenue" for 2027 would be $4.59T — far beyond what a $118T world GDP can plausibly support; the author notes actual revenue will be much lower.
The author genuinely engages both sides: Jevon's paradox proponents have a strong track record, but hard math across scenarios suggests overbuild risk is real, and the analysis is open for review.
More from Venture
- Windfall 2.5 ships recurring-payment tracking for its no-bank-link, local-model expense tracker — rudrank · 2026-11-01
- Prediction: Next year is the year of AI-run autonomous online businesses — curious_vii · 2026-11-01
- 1B white-collar workers × $200/month: why AI revenue may be a $2.4T market, not a bubble — simmol · 2026-11-01
- Analyst maps OpenAI vs Anthropic coding race: Opus 5.5 sparks Anthropic revenue rebound — firstadopter · 2026-11-01
- teortaxesTex suspects journalists and finance bros are using FUD to cool the market — teortaxesTex · 2026-11-01
- Intel crashes 30% in worst day in 50 years as Anthropic junk-rating rumor spreads — teortaxesTex · 2026-11-01