Token Revenue Are Growing. But That Doesn’t Mean We Are Not in a Bubble.
URL: https://hanchenli.github.io/blog/posts/token_economy/
Author: Hanchen Li
Published: 2026
Summary
Revenue figures for AI companies are staggering (Anthropic: 20B+ in 2025, Cursor: $2B ARR in 24 months). But this author argues much of today’s token spending is “exploration demand” — companies overspending to “not miss the AI revolution” without validated ROI — analogous to the 1990s dark fiber overbuild.
Key Claims
- Revenue numbers are real but don’t prove sustainability. Revenue = spending, not value delivered. Two very different forces drive token spending:
- Validated demand: measurable business returns (cost reduction, revenue generation)
- Exploration demand: “we need to try this” spending without proven ROI
- Exploration demand evidence:
- 42% of companies scrapped most AI initiatives in 2025 (up from 17% prior year)
- Only 5% of AI pilot programs achieve rapid revenue acceleration (MIT, Aug 2025)
- Average sunk cost per abandoned AI project: $4.2M
- Dark fiber analogy: 1990s internet built vast broadband infrastructure before Netflix/YouTube existed. Technology was right, timing was wrong, investors went bankrupt. Tokens may be this generation’s broadband.
- Jevons Paradox trap: Lower token costs → more usage → apparent demand growth. But lower cost demand isn’t the same as willingness to pay for the value delivered.
- Jensen Huang’s quote (GTC 2026): “The future data center is a token factory.” Every engineer will need an annual token budget worth half their base pay.
- The key question: How much of today’s token demand is long-lasting vs exploration money that vanishes when economy tightens?
Tension with Alex’s Work
Alex’s work in AI product and agents is directly affected by this dynamic. Building real value with AI is hard to prove even when the product ships — Hanchen’s framing helps distinguish products that survive a tightened market from ones that were exploration-funded.