For most of the past three years, the operating assumption for AI buyers has been simple. Models get cheaper. Models get better. Usage scales on a curve that bends toward zero marginal cost. You could disagree about timing. You could not really disagree about direction.
That consensus broke this week. Three independent signals, from three people with no reason to coordinate, published inside seven days. Each attacks a different axis.
Act one: the cost of doing one unit of AI work is rising, not falling. Act two: even when efficiency improves, buyers spend more, not less. Act three: the revenue numbers that justify the spending are inflated by an accounting convention most boards have not examined.
The through-line is not pessimism. It is measurement. When headline numbers become unreliable on three sides of the same equation at the same time, the durable advantage shifts to whoever can tell the truth about what their AI actually did.
Act One: Toby Ord and the Rising Cost Per Hour
Toby Ord has modeled frontier AI economics for years. In April 2026, he asked a simple question. If token prices keep falling, why do agent tasks keep costing more?
His math is uncomfortable. Over the last seven years, frontier model parameters grew roughly 4,000x. Token consumption per task grew roughly 100,000x. The drop in cost per token did not keep up with the explosion in tokens per task. The result is a cost-per-hour curve that bends the wrong way.
Using public pricing, Ord places today’s frontier reasoning agents in a sweet-spot band of $0.40 to $40 per hour of equivalent work. Cheap, by any historical measure of professional labor. But push the same models to saturation, where they think harder on harder problems, and the range opens from $13 to $350 per hour. The human software engineer baseline he uses as a comparison sits near $120 per hour, fully loaded.
Read those ranges carefully. At the low end, AI is a genuine discount against human labor. At the high end, the frontier model costs almost three times the human. The high end is exactly where the hardest, most-valuable tasks live.
Ord’s summary: “These cutting-edge AI systems would be getting less cost-competitive with humans over time.” Not a forecast. A description of what the pricing already shows.
One caveat. The $350 figure is the saturation case, not the typical case. A typical agent on a reasonable reasoning budget looks much closer to the sweet-spot band. But the hardest enterprise problems, the ones most likely to justify AI spend in the first place, push toward saturation. Pricing a program by the sweet-spot case and deploying at saturation is how budgets get surprised.
Act Two: Cursor and the Jevons Signal
Three days later, Cursor published an internal study of usage across 500 companies over eight months. The headline, in their words: better models drive more usage, not less.
Between July 2025 and March 2026, weekly AI messages per engineer in their sample rose 44%. Not at companies that adopted Cursor that quarter. At companies already using it. Same seats, same engineers, 44% more messages.
The breakdown earns the piece its weight. High-complexity coding rose 68%. Documentation rose 62%. Architecture rose 52%. Code review rose 51%. Simple UI styling, the category you would expect to benefit most from better models, rose only 15%.
Cursor’s interpretation: “Better AI leads to greater AI demand. This is consistent with a Jevons-like effect, where gains in efficiency increase total consumption rather than reducing it.”
We have written about this pattern. In AI Economics Governance Moat, the argument was that cheaper tokens expand your governance burden rather than shrinking it. In The AI Intensity Trap, the warning was that “more usage” is not “more output.” Cursor’s data is the cleanest industrial evidence we have seen for both.
Caveat this study the way you would any vendor’s own data. Cursor users are AI-positive by selection. 500 companies that chose to pay for an AI coding tool is not the general market. But the shape of the finding is what matters. When capability expands, the hardest categories absorb the most incremental usage. The easy work barely moves. That is the opposite of what a “capability surplus that frees up humans” world would look like.
The budget consequence is blunt. Your AI bill did not go down. Your AI usage went up 44% because the models got good enough to let you try harder things. If leadership asked for cost savings and your engineers did what good engineers do with better tools, you did not get savings. You got ambition.
Act Three: Scott Stevenson and Contracted ARR
The third signal arrived via a Scott Stevenson post on X on April 17, surfaced in that morning’s TLDR. The claim is blunt. A growing share of the AI-native SaaS revenue figures being celebrated in the press are inflated by a convention called “contracted ARR.”
The mechanics, simplified. A vendor signs a three-year deal. Under a standard ARR definition, one year is reported as current ARR. Under “contracted ARR,” the full committed value across the contract window is reported as current revenue. On a single deal, a rounding error. On a fast-growing AI startup with a portfolio of multi-year commitments, the gap runs 2x to 3x. The lower number is what the business actually runs at. The higher number is what the press release prints.
Stevenson’s phrasing is sharper than ours: “Every VC has a fraud in their portfolio.” We would put it carefully. This is a disclosure practice, not a fraud. It is legal. It is increasingly common. And it makes the published growth rates of AI-native SaaS companies effectively uncomparable with the historic SaaS benchmarks being used to justify valuations.
Treat the signal as a practitioner’s claim, not audited data. One credible voice, not a regulator’s finding. But the direction matches what boards have been asking us about privately for six months. When the growth number cannot be tied to cash collected in the period, downstream planning built on that number is already wrong. The question is only when the correction arrives.
The Convergence
Any one of these signals would be a useful data point. What makes this week unusual is that all three arrived in the same window, attacking different faces of the same dome.
Ord fractures the cost assumption. Cursor fractures the efficiency-leads-to-savings assumption. Stevenson fractures the revenue-growth assumption. Together, the three numbers most often used to justify AI programs, cost per task, productivity gains, and vendor revenue growth, are each moving in ways the last two years of narrative did not prepare anyone for.
This is not a crash. The capability is real. The usage is real. The revenue, even at the honest ARR, is real. What changed is the confidence interval on every headline number.
In AI Value Chain Inverted Economics, we argued that concentration at the chip layer made enterprise AI strategy structurally fragile. This week adds a second fragility. It is not just that 79% of the profit pool belongs to one vendor. It is that the customer-side numbers used to defend AI budgets are themselves under pressure from three directions at once.
Measurement Integrity Is the Moat
When capability is a commodity, cost is rising, and revenue is contested, the advantage shifts to whoever can measure what actually happened. Not what was promised. Not what was booked. Not what the dashboard defaults to. What actually happened.
A governance frame makes this concrete. If Ord is directionally right, your AI cost model needs a saturation case, not just a sweet-spot case. If Cursor is directionally right, your efficiency gains belong in output per dollar, not tokens per minute. If Stevenson is directionally right, any vendor you rely on should be evaluated against current-period cash revenue, not contracted headline figures.
None of these are new disciplines. They are what a competent CFO applies to any category of spending. What is new is that AI is now large enough, and its reporting inconsistent enough, that treating it like any other line item is the differentiating move.
The firms still spending on AI in three years will not be the ones with the best model access. Model access is table stakes. They will be the ones whose internal measurement is honest enough that their own numbers survive contact with Ord’s cost curve, Cursor’s usage curve, and the audit Stevenson’s post will eventually provoke.
Measurement is not glamorous. It compounds. And in a week where three axes of the consensus broke at once, compounding honesty is what buys you a seat at the table after the corrections arrive.
This analysis synthesizes Toby Ord’s “Are the Costs of AI Agents Also Rising Exponentially?” (April 2026), Cursor’s “Better AI Models Enable More Ambitious Work” (April 2026), and Scott Stevenson’s “It’s Time to Expose Contracted ARR” (X, April 2026).
Victorino Group helps teams measure what AI actually delivers when headline numbers can’t be trusted. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation