Three numbers from this quarter’s filings:

1. $4.7 million. Reported revenue per employee for the largest US frontier lab, calculated against the most recent headcount disclosure. The same number, four years ago, was $360,000 — already a high mark by tech standards, unprecedented at scale.

2. $2.1 million. Reported revenue per employee at the second-largest frontier lab. The gap is closing slower than the comparative compute spend would predict. The two labs are operating at roughly comparable scale on the inputs and meaningfully diverging on the output ratio.

3. $185,000. Median revenue per employee at the cohort of mid-tier AI infrastructure providers two tiers down — the GPU brokers, the model-hosting platforms, the inference resellers. The number is healthy by any historic comparison. It is two orders of magnitude below the labs the cohort sells into.

The shape of the chart, if drawn properly, is not a slope. It is a step function. Two companies sit at one altitude; a small handful sit at a second altitude; the rest of the industry sits at a third. The altitudes are separated, in revenue-per-head terms, by orders of magnitude.

This produces a comp-ladder problem the broader market has not yet solved.

A US software engineer with five years of experience can, in 2026, command a base salary that varies by a factor of ten depending on which altitude they are working at. The engineer at the frontier lab is, in revenue-per-head terms, defensible at the top of that range. The engineer at the third-altitude infrastructure provider is, in the same terms, defensible at the bottom. The engineers in between are negotiating against a data set that does not have a clean middle.

The result, observable in compensation surveys this quarter, is bimodal. The middle of the ladder has thinned. Engineers either get top-tier offers or third-tier offers; the second tier is where the offers have become hardest to land. The candidates are not avoiding the second tier. The second tier has become harder to defend internally because it cannot, on its own revenue-per-head, justify the ladder it is being asked to compete on.

A few structural notes:

— The revenue-per-head ratio at the top is being driven by API pricing, not by headcount efficiency. The labs are not staffing thinner than they used to. They are pricing into a market that has expanded faster than headcount can. This is a temporary condition. Headcount will catch up. The ratio will compress.

— The compression has not started. It will likely start when one of the top-tier labs begins hiring aggressively against a specific second-tier function — most likely deployment engineering or applied research adjacent to enterprise sales. When that hiring opens, the second tier will feel it within a quarter.

— The infrastructure providers at the third altitude are, in absolute revenue terms, growing fastest. They are not growing fastest in revenue-per-head because their headcount is growing in lockstep. This is, for them, a strategic choice. They are buying coverage, not margin.

The takeaway is not that the labs are overpaying. It is that the comp ladder, structured against revenue-per-head, is producing salary outcomes that the market cannot rationalise across all three altitudes at once. The middle is being pulled in two directions, and the engineers in the middle are, predictably, the ones with the most leverage to move.

The market will sort this out. It will sort it out unevenly, and the unevenness will be the story of compensation in this industry for the next eighteen months. The number that broke the ladder is $4.7 million. The number that re-builds it will be whatever the second tier negotiates as their floor. That number is being set, now, in offer letters that will not be public for another two quarters.

When it is, the chart will redraw. Until it does, the bimodality is the data.