When Jeff Dean walked out of Google to start his own lab, the market did the math in a single day. Dean spent 27 years building the infrastructure that trained the world's models, and his exit, alongside four senior researchers and a chairmanship reshuffle for DeepMind's Demis Hassabis that looked a lot like a soft landing, knocked Google's stock down 4 percent, roughly 200 billion dollars of value. Investor Brad Gerstner of Altimeter Capital, speaking on the All-In podcast, argued this is being read wrong. It is not a personnel story, he said. It is a capital-allocation story, and the signal is blunt: the money is no longer on the science, it is on the steel, the concrete and the power lines.
The arithmetic behind that claim is worth spelling out. Google has committed to around 200 billion dollars of AI capital spending this year, most of it on data centres. Under current US tax rules, accelerated depreciation hands back about 26 cents on every dollar poured into domestic compute, a government-subsidised, high-confidence bet at a moment when demand for compute is off the charts and Google is one of the best infrastructure operators alive. Frontier model development is the opposite kind of wager: tens of billions per training run, with far less certain payoff. As David Friedberg put it on the same show, "capex is high alpha, low beta" in data centres, while model development is "a very risky way to deploy capital."
What has changed the calculus is competition from the open-weight world. Models out of China and the open-source community are closing the capability gap every quarter, compressing the premium that once rewarded whoever reached the frontier first. If a near-frontier model can be downloaded for free within months of a breakthrough, the economic value of being first shrinks, and with it the case for spending tens of billions to get there. Better, the logic runs, to own the tollbooth every model has to drive through than to bet on building the fastest car.
The talent is reading the same spreadsheet as the capital. A scientist of Dean's stature, Friedberg noted, can "raise a couple billion dollars at a multi-billion-dollar pre-money with a PowerPoint deck." So the researchers are flowing out of the big integrated labs and into independent frontier plays or toward infrastructure, following the money rather than the org chart. Gerstner added a structural wrinkle: every cloud owner now faces a channel conflict, because Google Cloud would rather rent its scarce GPUs to Anthropic, which pays top spot prices, than burn them on its own uncertain research.
The uncomfortable irony sits underneath all of it. This reallocation is happening at the exact moment the labs insist that scientific progress is accelerating, that models are starting to improve themselves. If the smart money is quietly deciding that discovery is too risky and infrastructure is the safe return, it is making a bet against the very breakthroughs the industry keeps promising. Someone still has to fund the science. The question Dean's departure raises is who, and at what price, once the capital has decided it would rather own the picks and shovels.