Dwarkesh Patel speaks with SemiAnalysis founder Dylan Patel about how AI labs’ economics and infrastructure demand could reshape technology markets and macro-finance. Patel claims the leading labs’ compute spending is accelerating, projecting that OpenAI and Anthropic could take 40–50% of incremental compute next year and roughly half by the end of the following year, with newer chips amplifying their share of “usable” performance. He also asserts that model-layer margins have swung sharply, saying Anthropic became profitable in Q2 and that OpenAI could follow in Q3, as revenue per megawatt rises.
The pair argue over physical and supply-chain bottlenecks, including constraints around EUV tooling and component capacity, and discuss how compute pricing could be bid up if frontier labs can monetise each additional megawatt better than rivals. Patel alleges regulatory and safety pressures are already slowing leading labs’ releases and training cadence, potentially narrowing their monetisation edge.
On geopolitics, Patel estimates China’s current share of incremental AI data-centre compute is below 10% while the US is around 70%, attributing the shift to export controls and US capital deployment. He predicts China’s build-out could inflect later in the decade as domestic fabs ramp, though with lower-performing chips.
They close on a macro thesis: if AI infrastructure requires trillions in capex funded partly by debt, competition for credit could lift spreads and rates, squeezing other sectors and increasing default risk for highly indebted countries.
Dylan Patel, founder of SemiAnalysis, says Anthropic became profitable in Q2 and that OpenAI could turn profitable in Q3, attributing the shift to rising model-driven demand and improving unit economics. He frames the change as a recent break from what he describes as “venture-funded losses” dominating lab finances a year earlier.
Read this moment →Patel claims that safety and regulatory pressure is constraining frontier labs’ deployment and training tempo, and argues the effect is asymmetric versus open-source Chinese models. He cites, as examples, OpenAI “not releasing Astra” and “stopping training for two weeks”, and also alleges that labs are not releasing their best models to the public, which would affect revenue per unit of compute.
Read this moment →