The lab-to-lab talent market is too small

There is a default hiring strategy at the frontier, and almost every organization runs it. Identify the people doing comparable work at peer labs, approach them, and compete on compensation, compute, and mission. It is a rational strategy for any individual hire. It is a poor strategy for an industry, because the population it targets is small, fully employed, approached constantly, and growing far more slowly than the demand for it.

The argument of this briefing is that the lab-to-lab market is now structurally insufficient, not merely competitive, and that the teams pulling ahead on hiring are the ones that have stopped treating an existing frontier-lab title as the primary qualification.

Consider the supply side honestly. The number of people who have worked on a large-scale training run, or owned a serving system under real load, or built an evaluation harness that survived contact with a strong model, is measured in the low thousands globally and plausibly in the hundreds for the most specialized functions. That population grows only as fast as labs can train people inside it, which is to say slowly, and each departure is another lab's gain rather than a net addition.

Meanwhile demand has expanded along two axes at once. More organizations are attempting frontier-adjacent work, and each of them needs a wider set of functions than it did two years ago, as the previous briefing argued. A market where demand grows on two axes and supply grows on none is not a competitive market. It is a shortage, and shortages are not solved by better outreach to the same names.

The observable consequence is compensation escalation without a corresponding increase in hires made. Every lab reports difficulty filling the same roles. That is the signature of a supply problem being addressed as though it were a persuasion problem.

Broadening the pool is only useful if the adjacent markets genuinely produce the capability required. In several cases they demonstrably do, because the underlying work is the same and only the domain differs.

Quantitative research and trading. The daily rhythm is designing an experiment, building the system to run it, measuring the result carefully, and discarding most of what was tried. That is the research-engineering loop with different subject matter. These candidates also arrive with a default suspicion of results that look too good, which is a substantial asset in a field where measurement is genuinely hard.

High-performance and scientific computing. Physics, computational biology, astronomy, and climate modeling produce people who have spent years building large-scale simulation and analysis systems to answer questions nobody knew the answer to. They have written a great deal of scientific code and are comfortable with a moving target. What they lack is current machine learning vocabulary, which is the most learnable part of the requirement.

Distributed systems and developer infrastructure. Engineers who have operated large fleets understand failure as a statistical property rather than an incident, which is precisely the mental model a large training run demands. Those from developer-tooling backgrounds bring something rarer: the instinct to treat other engineers as the customer, which is what research tooling actually is.

Robotics and safety-critical engineering. Autonomous systems, aerospace, and medical devices train people to treat measurement and failure analysis as first-class engineering problems, under conditions where being wrong is unacceptable. That habit transfers directly to the evaluation-heavy end of the work, which is where several labs are currently thinnest.

Widening the aperture fails if the assessment stays the same, and most assessments quietly encode the old pool. A screen built around familiarity with current architectures will reject a computational physicist who could have closed that gap in a month, while passing a candidate who has the vocabulary and has never designed an experiment worth running. The screen is measuring exposure, not capability, and the two have diverged.

Two adjustments do most of the work. The first is to assess on the actual work: hand the candidate a real question, badly specified, and watch how they turn it into something measurable. This separates temperament from vocabulary in a way no knowledge screen can. The second is to decide internally, before the search opens, which version of the role is being hired, because titles at the frontier span work that is not interchangeable.

There is a timing argument as well. Candidates from adjacent markets are less approached, more receptive to a first conversation, and weighing a genuinely interesting change rather than a lateral one. They are also slower to move, because the decision is larger and less familiar. A search that has mapped these people in advance can afford those conversations. A search that opened last week cannot, which is exactly why it ends up back in the contested pool with everyone else.

The labs staffing these roles well are not, for the most part, winning on money. They are willing to recognize the profile before the market has labelled it, and they have built a process that can evaluate someone who arrives without the expected credential. That is a hiring capability, and it is currently scarcer than the talent it finds.