When it comes to attracting AI talent, some companies spare no expense, but Rilla has raised the bar in a hard-to-ignore way. The startup, which builds AI-powered coaching software for sales teams, spends around $1.7 million a year on housing stipends for its employees, with a strict condition: they must live within a ten-minute bike ride from its New York office. The trade-off is a 72-hour workweek.

The arrangement, reported by Fortune, is not just another Silicon Valley-style perk. It signals an AI labor market so fiercely competitive for specialized skills that a company is willing to take on a housing cost normally borne by workers, just to eliminate commute time and maximize in-office presence.

A sign of structural stress

Behind the dollar figure is more than a well-funded startup’s indulgence. AI is undergoing a phase of acceleration that puts intense pressure on development teams. Shipping features ahead of competitors, training models on fresh data, iterating rapidly on customer feedback — all of this demands compressed work cycles. The 72-hour week is no accident: it’s a deliberate choice to squeeze development timelines, typical of a sector where time-to-market is everything.

But this choice has second-order implications that extend beyond Rilla’s walls. The “we pay your rent, you stay close” model tilts the playing field toward companies with enough capital to cover steep fixed costs, further concentrating talent in already hyper-expensive cities like New York or San Francisco. Top professionals get drawn in by packages that include housing, making it harder for smaller or more distantly located firms to access the same skills.

For those building AI in on-premise contexts or far from tech hubs, the dynamic is worrying. If the best engineers and data scientists cluster within a few square kilometers of Manhattan, companies wanting to deploy Large Language Models locally — perhaps in Europe or less central regions — will find an even tighter and more expensive labor market. And that’s without factoring in turnover: working 72-hour weeks for extended periods is unsustainable for many, and the risk of burnout could wipe out the initial investment in housing incentives, creating a hire-and-quit cycle that destroys institutional knowledge.

Winners and losers

In the short term, the winners are startups that can afford to buy time and proximity. They can iterate faster, test hypotheses, and bring products to market at record speed. The losers are the distributed-work models that many hoped would become the norm after the pandemic. Rilla’s is a bet on physical co-location as a productivity accelerator, and if other AI firms follow suit we’ll see further overheating of real estate prices and salaries in tech hubs.

Structurally, this story signals that AI is drifting away from any narrative of work democratization. The resources needed to compete — capital, hardware access, but also the ability to attract talent willing to put in grueling hours — are concentrating in few hands. It’s a wake-up call for anyone imagining a diffuse AI ecosystem built on on-premise deployment, edge computing, and local development: without policies to incentivize talent outside major hubs, the gap risks widening, not shrinking.