The Price Was Never the Point

The signal the market reads as a demand verdict is a volatile, window-dependent price. Whether the path moved with it is a fact the businesses already know and will not report until the second quarter.

The price is the residue of a cost that is falling on purpose. As I argued in The Collapsing Cost of Intelligence,² permanent price resets, distillation, and open and closed providers competing on one axis are driving the cost of a token down, and price falls with cost whether demand is weak or roaring. A cheaper token is what an efficiency gain looks like, not a statement about how many tokens anyone wants. The right question, which I raised in Seats Were Always a Proxy,³ is whether units are being consumed in rising volume, by more customers, for work that matters. That is the token path, and it sits one layer in from the price.

The consumption itself is not in doubt. There is a host of anecdotal evidence that consumption is still accelerating. But it is directional, not conclusive, because optimization keeps changing what a token is: efficiency compresses the tokens a task needs while reasoning and agents inflate them. And there is a trap. If a provider is paid by the token and each unit of value keeps getting cheaper to produce, revenue can deflate even as usage climbs, unless it captures value beyond the token itself. No token counter scores that. The truth arrives, ultimately, on the income statement.

Read the path with two tests, not with a price screen. The first is the treadmill test: does revenue rise with the value consumed, even as the token cost of delivering it falls? A business that passes is monetizing consumption; one that fails is repricing an old license. The second is the routing-discipline test: as cheaper models arrive, does the provider route work down to them and keep the margin, or get commoditized as the price it charges chases the price it pays? Neither shows up on a price screen.

The tests point at a particular kind of company. The businesses that pass are the rails-owners I described in Seats Were Always a Proxy:³ those that sit on the process flow, compound proprietary data, and carry the compliance depth that makes their pipes hard to replace. For them a cheaper token is a tailwind, widening the base of work worth routing through pipes they are paid on. The toll-taker who only resells model output just watches its cut thin.

And reported earnings lag the operations they describe. The consumption names I can see grew revenue through April, but the index did not peak until late May or fall until summer, so what we can observe is the climb, not the drop. So far, so good. The quarter that matters has not printed. It will be a real interim test, in retention and remaining performance obligations, but not likely the final answer.

Reading across: the IBM warning

As this piece went to press, a large legacy vendor pre-announced a weak quarter and lost about a quarter of its value in a day, its worst on record,⁴ as clients shifted budget from software toward AI hardware and a memory-price spike crowded out the rest. Annual IT budgets were set last year, before the hardware scramble, so a fixed pool is being redirected mid-cycle and software becomes the bill-payer. Peers fell in sympathy, consumption platforms among them, on the reflex that one vendor’s miss is the sector’s tell.

Read across with care. A heavy services and legacy mix is where AI substitutes, not where it compounds, and it is not the franchise. The company still expects roughly $15.7 billion of free cash flow this year, enough to keep raising a dividend lifted for more than three decades and to underwrite a quantum roadmap aimed at a fault-tolerant machine before the decade is out^5,6. A budget vintage caught offside is deferral, not disappearance, and the read that matters is each company’s own results.

The gap is where the opportunity sits, and there are two of them. In Capturing Pent-Up Alpha⁷ I described the later one: the return that accrues when reported results and market price come apart, a gap the market is slow to close for structural reasons, not because the evidence is hidden.6 There is an earlier gap upstream of it, between what a business is doing and what its numbers have yet shown. Call it the pent-up inflection, the backward extension of the same idea, and it is where the token path sits today: the turn is visible in consumption but has not reached the income statement, so the alpha has not begun to accrue.

None of this says token prices will stop falling; a falling cost of intelligence is a feature of a healthy cycle, not a warning. The signal the market thinks it lost was never the one it needed. The price was never the point. The path was, and for now it is still yours to read before the market catches up to it.

Footnotes 

¹Jan-Patrick Barnert and Michael Msika, “With token prices collapsing and regulation rising, AI’s pricing power looks fragile”, Los Angeles Times, July 3, 2026.
² Mark Scalzo, “The Collapsing Cost of Intelligence“, validex.co, June 30, 2026.
³ Mark Scalzo, “Seats Were Always A Proxy“, validex.co, May 19, 2026.
⁴ Yun Li and CJ Haddad, “IBM stock craters 25%, the worst day on record, after company issues second-quarter earnings warning”, CNBC, July 14, 2026.
⁵ International Business Machines Corporation, “IBM Releases First-Quarter Results,” Exhibit 99.1, Form 8-K, filed with the U.S. Securities and Exchange Commission, April 22, 2026.
⁶ MLQ.ai, “IBM Projects $15.7B Free Cash Flow for 2026 Amid Confidence,” IBM Q4 2025 Earnings Highlights, (accessed July 16, 2026).
⁷ Mark Scalzo, “Capturing Pent-Up Alpha“, validex.co, March 30, 2026.