Token costs force CFO choice: pay the inference bill or keep headcount

CNBC reports CFOs now face a binary choice: tokens or humans. AI inference costs are running ahead of enterprise forecasts, and the overage is not getting absorbed by productivity gains. The rack math was always going to land here. Model providers priced inference assuming marginal-cost decline that has not materialized at the rate promised, and enterprises budgeted assuming token consumption would plateau. Neither happened.

Glean crossed $300 million ARR by positioning as a cost-cutting solution, tripling revenue in twelve months. That growth rate in the current financing environment tells you where enterprise AI budget is flowing: toward vendors who promise to reduce total spend, not vendors who promise to unlock new capability. The substitution dynamic is live. Enterprises are comparing token cost per task against fully loaded FTE cost per task, and in a tightening capital window the comparison is not theoretical.

Anthropic's valuation hit near-trillion-dollar territory after its latest raise, and the seven cofounders now hold $116 billion in combined paper net worth. That valuation assumes sustained pricing power and margin expansion at the model layer. The enterprise trade-off CNBC describes is the counterparty risk to that assumption. If token costs compress headcount budgets and CFOs respond by rationing inference volume, the revenue model underneath the frontier providers reprice downward or the volume assumptions in their financing structures do not clear.

The current spread between sticker API pricing and effective cost per task is wide enough that some enterprises are running pilot programs to quantify the trade-off with three significant figures. When those pilots finish and procurement starts negotiating on a cost-per-replaced-FTE basis rather than cost-per-million-tokens, the model providers will face margin compression from the demand side, not just the supply side. The open-weight models are not yet fast enough or cheap enough to substitute at enterprise scale, but the substitution threshold is now defined by a specific dollar figure in a specific CFO's budget deck.

Glean's revenue trajectory and Anthropic's valuation are both real. The tension between them is also real. One is built on enterprises cutting costs. The other is built on enterprises paying the inference bill at current rates. Both cannot be true at the same time for the same customer.