The capex curve is flattening. Not declining — flattening. Two charts make the point.

Chart one: the year-over-year growth in compute capex at the top three frontier labs, plotted from 2022 to Q1 2026. The shape is a steep linear rise from 2022 through 2024, a slight inflection in 2025, and a noticeable softening in the first quarter of this year. The 2026 trajectory, if extrapolated from Q1, lands at roughly 8% growth — meaningful, but a fraction of the 65% growth seen in 2024.

Chart two: the same three labs’ year-over-year growth in research headcount, over the same window. The shape is closer to a step function. A flat baseline through 2022, two steep step-ups through 2024, and a third step-up partway through 2025 that has held flat into 2026.

The compute spend is decelerating. The headcount is not.

The implication is one the industry has been uncomfortable saying out loud. The next foundation lab is going to look, in capital structure, less like a research lab and more like a trading desk: moderate fixed compute outlay, large pool of high-leverage individual contributors, capex amortised over a long run rather than torched in a single training cycle. The economics that made the 2022–2024 lab work — buy more GPUs, train a bigger model, ship it, bank the lift — is no longer the economics that produces the marginal performance gain.

This is the architectural answer to a question the industry has been asking in financial language. Why has the marginal dollar of compute stopped producing what it used to. The answer, increasingly, is that the marginal dollar of compute hits a ceiling that the marginal dollar of research talent does not. The training run that justified the 2024 capex is not the training run that justifies the 2026 capex.

A few structural observations:

— The chip-supply situation, which drove most of 2022–2024 capex, is no longer a binding constraint at the very top of the market. The constraint has migrated to model-architecture innovation, which is talent-bound rather than capital-bound.

— The largest labs are, accordingly, reorganising. Two of them have publicly restructured their research divisions in the last six months, in ways that resemble more closely the desk-and-PM structure of an investment bank than the principal-investigator structure of an academic lab. This is not coincidence.

— The labs that have not made this transition are, in Q1 reporting, showing the slowest performance gains. The capex spend at those labs is unchanged. The output-per-capex ratio has fallen.

The next round of foundation labs — the ones launching this year and next — will, almost without exception, launch with this thesis. Less compute. More compensation. A capital structure that resembles a trading floor’s. The pitch decks already read this way.

This will change what the labs look like in public. The aesthetic of the lab — campus, whiteboards, papers — will give way, slowly, to the aesthetic of the desk: floor, screens, allocations. Both can produce frontier work. They produce different cultures while doing it.

The integrative read is structural, not normative. The capex curve is flattening because the constraint is moving. The institutions are reshaping around the new constraint. The new shape will look familiar to anyone who has spent time in financial markets and unfamiliar to anyone who has spent time in academic research. Both views are correct. The lab is becoming, in part, a desk.

The chart is the data. The reorganisation is the response. The next two years are the working-out of the response.