Brad Gerstner: Google's AI Brain Drain, SpaceX's Huge Quarter, Airtable's 90% Collapse, US Data Fuels China AI

This episode opens on August 6 with the show short a host. Chamath Palihapitiya is travelling, so Brad Gerstner fills the seat, and Jason Calacanis spends the first minutes converting that absence into material: a photograph presented as a field report from a data centre site, a joke that the location is actually the set of Dune, and a running gag about David Sacks arriving late because important people keep calling him. Sacks eventually appears in what he calls a summer gelee, and the hosts spend another beat on the shirt before conceding they have real material to get through. The banter is not filler so much as a statement of the show's register: four investors who will, within twenty minutes, be pricing gigawatts of compute and arguing about whether a frontier model is a defensible business.

The first substantive story is a pair of personnel changes at Google on the Wednesday before recording. Demis Hassabis has moved to chair of DeepMind and chief scientist at Google. The hosts note the two available readings immediately: either he stepped down, or he was kicked upstairs. Google's own framing is that this is a promotion and that he is stepping up. What makes the item newsworthy is not the title change but the reporting attached to it. Axios, quoted on the show, says Gemini 3.5 Pro is months behind, and that company sources attribute this in part to low morale. Several top researchers, including a co-lead of Gemini, have left for competing labs.

The larger departure is Jeff Dean, who is leaving with three other AI superstars to start a company called Discovery Loop, focused on deep scientific breakthroughs in artificial intelligence. Gerstner supplies the biographical weight: employee number thirty, joined in 1999, and by his understanding continuously employed there for twenty-seven years. He also supplies the market's verdict, noting Google shares fell 4% on the news, which he translates into roughly $200 billion of lost market capitalisation if the two events are correlated.

Friedberg is asked whether this is creative destruction, a failure of delivery that justified fresh blood, or simply the pull of startup opportunity in an era of unlimited AI capital. He rejects both framings for a third, and it is the analytical spine of the whole segment.

Friedberg's argument begins with a boardroom debate about capital allocation rather than a talent story. Google has committed to deploy $200 billion of capex this year on AI infrastructure and data centre build-out. Two features make that spending unusually attractive right now. Accelerated depreciation means an investment in US AI compute is heavily tax-advantaged; at a 26% corporate rate, Gerstner later adds, a dollar of capex is effectively 26% off because it can be written off in the same year. And demand for compute is so extreme that the return on invested capital is close to obvious. If you are good at operating compute infrastructure, that capital can produce large profit returns with very high confidence inside a forecastable period.

A frontier model, by contrast, also costs tens of billions, but the question of whether you can deliver profit from it is genuinely open. Friedberg points to three simultaneous compressions: open source is getting very good, open-weight models are catching up quickly, and the frontier labs are converging on each other. In that world, the marginal dollar is better spent on the layer where demand is contractual and the asset is model-agnostic.

He then reframes the departing scientists. These are the people who were at the core of model development, and they were first out of the gate: early interviews with Dean describe an internal ChatGPT equivalent a year before OpenAI shipped one, which Google declined to release for fear of cannibalising search. Sergey Brin's return produced the revitalisation. But as competition on models intensified, the returns to model capital stopped looking like the returns to infrastructure capital. Google's real asset is distribution: the largest enterprise install base and the largest consumer base, neither of which requires owning the best model. It can host Anthropic, OpenAI, and open-weight models, and it already holds significant stakes in SpaceX and Anthropic.

The mechanism that turns this into departures is internal. If capital is flowing to data centres and to a broad model ecosystem rather than to your model, and if you can walk down the street and raise a couple of billion at a multi-billion pre-money on the strength of a deck, the outside option dominates. Friedberg's formulation is deliberately financial: data-centre capex is high alpha and low beta; model development may be high alpha but it is very high beta. A board maximising risk-adjusted return deploys more into compute and less into models. That, he says, is what is going on.

Gerstner endorses the analysis and generalises it. Microsoft is running the same play: Satya Nadella, citing a Morgan Stanley report, describes returns above 30% on invested capital in tokens-as-a-service, meaning the infrastructure business. The scientists who want to work on superintelligence, cure cancer, or sit at the model frontier are instead navigating a channel conflict. Google Cloud wants the compute so it can rent it to Anthropic; the internal frontier teams want the same compute so they can compete with Anthropic. The conflict appears to be resolving in favour of being an infrastructure company.