AI tool spending in engineering will rise tenfold in share, then fall in absolute terms
Matteo Franceschetti makes a prediction that accepts a rising usage curve and bets against a rising cost curve at the same time. The call is specific enough to be wrong, which is what makes it worth examining.
Most cost projections for AI tooling in engineering teams assume the trend runs in one direction. Matteo Franceschetti disagrees with the conclusion while accepting the premise. His call: usage as a share of engineering salary will climb from five percent to 50 percent, and the total bill will still shrink. The bet lands or fails on a single variable, the rate at which model costs fall relative to volume growth.
The structure of the argument deserves precise statement, because it is easy to blur. Franceschetti is not predicting that AI tools get cheaper in isolation. He is predicting that deflation in model costs will outpace the expansion in consumption volume fast enough to produce a lower absolute spend figure. That requires the cost curve to move more steeply downward than the usage curve moves upward. Both are moving quickly, which is what makes the call falsifiable rather than vague.
The counterargument is equally structural. If the productivity gains that justify heavy AI usage fail to materialize across a broad enough range of engineering roles, companies face pressure to constrain volume before adoption reaches the 50 percent figure. That caps usage without producing the cost savings either. Franceschetti’s net-spend-falls prediction requires both high adoption and steep deflation simultaneously. Either condition slipping changes the arithmetic in ways that the headline call obscures.
The usage will increase and so the 5% will become 50%, but the cost will go down and so net that I think it will go down. Matteo Franceschetti
There is also a near-term dynamic that runs against Franceschetti’s endpoint. If AI tool costs as a share of engineering salary are still climbing now, companies calibrating to current spend levels will be setting budgets against a curve that has not yet bent. The reconciliation of a rising near-term trajectory with a falling long-term one depends entirely on timing, specifically on when deflation becomes steep enough to reverse the direction of the total line. Franceschetti’s call does not pin a year to that inflection, which is a genuine limitation for anyone who wants to track it against outcomes.
What can be tracked is the gap between usage share and net spend. If AI tool costs as a percentage of engineering salaries climb and then decline, the call is on course. If they climb and stay elevated, the bet is wrong. The observable variable is cost per unit of model output multiplied by total consumption volume, measured against engineering salary lines. That data exists at the company level, even if it is rarely disclosed publicly.
The mechanism Franceschetti gestures at has at least directional plausibility. Model pricing has generally moved downward as competition among providers has intensified and as infrastructure efficiency has improved. What Franceschetti is asserting is that this downward pressure is not a temporary feature of an immature market but a durable force strong enough to win a race against a tenfold increase in usage share. That is a strong claim about relative rates, not merely about direction. It implies that the deflationary force compounds over time rather than moderating as the market matures, which is not guaranteed by the direction of travel alone.
What rides on the call being right extends beyond corporate budget lines. If Franceschetti is correct, the period of visible AI spend growth in engineering is temporary, and organizations treating current cost levels as a durable ceiling will be calibrating to the wrong baseline. If he is wrong, and volume growth outpaces deflation, the economics of software teams will look structurally different within a few years, with AI tooling becoming a dominant cost line rather than a productivity supplement. The call is checkable. The ratio of deflation rate to usage growth rate is the number to watch, and any company running AI-assisted engineering at scale has enough internal data to know, earlier than the market does, which direction that ratio is moving.