Skill adjacency is doing more work in these answers than compensation is. A frontier lab buys people who make models smarter, and the asset held by a senior applied scientist on recommendations is the demand side of a marketplace, which by Abhinav Bohra's account the labs do not yet have: no catalog, no sellers, no shoppers to rank for [4]. The condition he sets is specific rather than sentimental, since he says the day a lab runs an actual marketplace, people like him get very interested [4]. That is a conditional poaching risk rather than a continuous one, and the trigger is a product decision at the lab, not a comp decision at Amazon.
The sample deserves stating plainly. Business Insider spoke with eight workers [1], and the material in front of us carries three of those responses, which leaves five we have not read [15]. Eight people is a thin sample, and three is thinner still, so this is not a rate. What a panel like this yields is reasons, and reasons are testable against an attrition file: if leavers cluster in model research while ranking and serving teams hold, the mechanism described here is the one running in your org. Of the three responses we can read, two describe declining to pursue lab jobs and the third names Anthropic as a dream employer without any application recorded [16].
Widening comp bands for AI roles and moving on treats this as a pay problem, but money defends against money, and neither refusal here is denominated in dollars. One is scope of work [4]. The other is balance-sheet risk: a Meta software engineer told Business Insider he has not applied to the labs partly because they run on investor capital while hyperscalers fund their AI push from core businesses that already make enormous money [11]. He also argues the hyperscalers are doing the same caliber of work, including training and the infrastructure to serve it to billions [10]. Against reasoning like that, a raise is beside the point, and after years of tech layoffs made job security a heavier factor [13], the argument is likely to be widely held.
The values case has a stated expiry, and the person making it says so. Mike Kostersitz of Nike names Anthropic as his choice outside his employer [6], citing its refusal of a Pentagon push to strip restrictions on mass surveillance and autonomous weapons use, which its own CFO said could cost billions in revenue [7]. He then adds that he does not expect it to hold, because an IPO pulls toward shareholder value [8]. A recruiting advantage its admirers already date is one a competitor can wait out.
This same evidence has a second reader, and it is Nvidia. Bohra ranks Nvidia behind only his current employer, on the argument that recommendation at scale is an inference cost problem, where running a model once is cheap and running it for every shopper in real time is where the money goes [5]. So the population that Big Tech must defend this quarter is narrow and comparatively cheap to hold. It widens the moment a lab opens marketplace work or starts paying for compute out of operating revenue, and both stated reasons for staying would go at once.