Everyone "knows" AI is sold below cost — but the subsidy is hiding in the opposite place from where you think. The metered API that charges by the token is profitable: OpenAI's Sam Altman says so, Anthropic's Dario Amodei puts inference margins above 50%, and teardowns put them higher still. The real giveaway is the flat $200-a-month subscription — a maxed ChatGPT Pro or Claude Max plan can burn 40 to 70 times its price in compute. This episode follows the money through both lies the price sheet tells: that cheaper-per-token means cheaper (it doesn't, once reasoning models' hidden "thinking" tokens are counted), and that the meter is where the loss lives (it isn't).

The deeper story is who survives the coming price war — and it's settled one layer below the price sheet, at the silicon. Nearly every lab rents NVIDIA chips and pays its ~75% margin as a tax on every token; Google escaped it by building its own TPUs, which is why Gemini can undercut OpenAI and Anthropic and still clear its cost — and why Anthropic is paying Google tens of billions to rent those same chips. We map where Google, OpenAI, Anthropic, DeepSeek, xAI, Meta and Amazon really sit, why falling prices and record losses coexist (the Jevons paradox), what a token's "true cost" actually is, and when your flat plan gets a meter.

I'm Dan. AI moves too fast to keep up with, so I built my own stack of AI tools to research, analyse, verify and illustrate the questions I can't stop thinking about — mostly to learn it myself, and I share what I find. AI-assisted, fact-checked, worth a second look.

Follow Dan's AI Intel in your podcast app so new episodes find you.