AI pricing will collapse to token cost, argues dev pushing back on Paul Graham
bendee983 · x · 2026-09-16
- Paul Graham argued that paying per token isn't the true unit of inference, since better models solve more per token, and someone should define a unit via a chain of progressively harder problems.
- bendee983 counters that value-based pricing can't stick — like pricing electricity by use case — and AI pricing will inevitably collapse to the cost of generating tokens plus a fraction of fixed costs, as competitors undercut any margin above cost.
More from Venture
- Anthropic's run-rate revenue surges 7x in months: $9B to over $65B by July 2026 — rohanpaul_ai · 2026-09-16
- Anthropic run-rate tops $65B by July 2026; Amodei won't bet on $100B extrapolation — rohanpaul_ai · 2026-09-16
- PE-shaped GTM made Dynatrace $2B in revenue but skipped the developers entirely — thedealdirector · 2026-09-16
- AI Perez: a Perez-framework theory of where AI value accrues as models commoditize — cpaik · 2026-09-16
- Token Compass raises pre-seed led by True Ventures to measure enterprise AI token usage — ivan_bezdomny · 2026-09-16
- AI coding startup Factory AI raises at a $5B valuation round — annbordetsky · 2026-09-16