Ways to Think About Token Pricing

Primary-fetched from ben-evans.com, 2026-07-09.

Key Contributions

• The core question: will foundation models become low-margin commodity infrastructure (value captured elsewhere in the stack), or will they retain pricing power? Evans frames this as unresolved, not answered.
• Two analytical approaches:
  1. Bottom-up — token price sits somewhere between sellers' marginal cost and buyers' ROI, but neither variable is currently knowable with confidence.
  2. Top-down — look at how comparable technology transitions played out and identify the building blocks / where value accrued.
• Current margin claim: inference runs at 40–50% gross margins, but this figure excludes training costs, "which is currently far larger than revenue" — i.e. reported inference profitability is not lab profitability.
• Supply crunch framing: "we're in a supply crunch, and this is unstable. All of the variables are in play." The first-half-2025 demand surge concentrated in one use case (software development) — a narrow base to extrapolate from.
• Demand-side uncertainty: ROI of the recent usage surge is unclear — "it's unclear how much of the surge... has an ROI."
• Four questions Evans says determine the outcome:
  1. Frontier adoption — how many use cases need a frontier (expensive) model vs. a cheap one?
  2. Frontier progress — can capability gains outrun efficiency/pricing pressure?
  3. Competitive dynamics — does competition intensify, or does the market concentrate?
  4. Value capture — does value sit in the model layer, or in the software wrapped around it?

Historical Comparisons

• Mobile data (closest analogy): cellular data traffic rose "several orders of magnitude" into a ~$1 trillion annual revenue industry on ~$200 billion of capex — yet "the stocks have gone nowhere." Value was captured upstream (devices, apps), not by the carriers who built the pipes.
• Semiconductors: TSMC holds a near-monopoly on frontier logic manufacturing, yet its net income last year was "$53bn, less than half of Apple alone" — monopoly position did not translate to majority value capture.
• Fiber optics: overbuild parallels exist, but fiber build-out was capped by physical construction constraints; AI compute capacity remains "far behind demand" by contrast, which weakens the direct read-across.

Theoretical Uncertainty

Evans stresses we lack a "good theoretical understanding of why these models work so well" — unlike 1995 (internet protocols) or 2010 (mobile), the underlying physics/economics of the technology are not settled, which limits how far any forecast (including his own) should be trusted.

Bottom Line

"For a different outcome, something needs to happen that we don't see yet" — i.e. absent a new network effect, a frontier breakthrough that outruns efficiency gains, regulatory capture, or a geopolitical shift, Evans' base case is commodity-infrastructure economics for foundation models.

Companies / Entities Referenced

Anthropic, OpenAI, Meta, Microsoft, Google, Apple, TSMC, Samsung, SK Hynix.

Source: Ways to Think About Token Pricing by Benedict Evans, published 2026-07-09.