subject:Following up on ep. 184's cost-per-token frame — from the infra side
Bentley, caught your NVIDIA deep-dive last month. The piece you flagged on cost-per-token as the real margin story is the exact curve we've been watching at Tidal Compute.
Context on me: I run infra at Tidal Compute (we just closed Series A). The reason the token-cost question matters more than people think: inference is the first line item that gets scrutinized when unit economics stop hiding behind growth.
I think there's a 40-minute episode in the delta between what models cost to train and what they cost to serve — and why those curves diverge. Happy to bring real numbers from our H100 cluster and the last six months of pricing data.
Either way — the show's been a regular drive-home companion for the last year. Thanks for making it.
— Theo