Nine-Figure AI Offers and Those Who Turn Them Down

The compensation packages offered to a small number of senior researchers have become a subject of public discussion, with figures that would have been extraordinary for a chief executive a decade ago being offered to individual contributors.

A few thousand people worldwide can lead a frontier run

The number of people who can lead a frontier training run is small. The cost of a failed run — in money and in time — is enormous. And the marginal value of a researcher who improves a model’s efficiency by a few percentage points can exceed the annual compute budget of a small company.

• Scarcity: the qualified population is in the low thousands worldwide.
• Leverage: one person’s decision can change a company’s roadmap.
• Defence: a company paying to keep someone is also paying to deny them to a rival.
• Signalling: high-profile hires affect investor confidence and recruiting.

Compensated like executives, collaborating like rivals

The obvious casualty is research culture. When individual contributors are compensated like executives, internal collaboration becomes harder and publication becomes riskier, because knowledge is a competitive asset.

The subtler casualty is the pipeline. Money that flows to a few dozen senior people does not build the next generation, and the fields with the strongest academic programmes — often publicly funded — are the ones seeing their graduates leave for industry.

Compute, freedom to publish, colleagues nearby

The interesting minority are the researchers who decline large offers. Their stated reasons are consistent: access to compute, freedom to publish, proximity to colleagues, and the ability to work on problems that are not tied to a release schedule. Those are the things money struggles to buy, and their scarcity is the best evidence that the market is not purely financial.

Multi-year vesting turns a salary into a lock

The packages used to retain senior researchers are often structured as multi-year vesting schedules tied to continued employment, which converts a compensation decision into a lock-in mechanism. The practical effect is to make leaving expensive at precisely the moment a competitor is most interested.

Companies have also used compute allocation as an inducement, offering a personal research budget measured in accelerators. For a researcher whose work is limited by access to hardware, that can be worth more than the cash.

Peer review happens later, or not at all

The most consequential casualty is publication. A field that historically advanced by sharing results now has a large share of its most important work conducted privately, and the peer review that used to catch errors happens later or not at all.

Both effects — concentration of talent and retreat from publication — push in the same direction: fewer independent checks on the claims that shape public understanding of the technology.