renting a gpu costs $35–50k a year. buying one costs $30k and it runs for up to 10 years. speechifyai did the math in 2022 and never rented again

cliff weitzman, ceo of speechify – ai voice and text-to-speech platform with 60m users – breaks down why owning gpus changed everything for them:

the economics were only part of it – when engineers rented gpus, they self-censored, afraid of burning tens of thousands per experiment. weitzman calls it the michael jordan problem: you want to train for the nba but you're paying $20 an hour for court time, so you need a hoop in your house.

speechify bought its own racks and each engineer now runs 5 to 18 agents on dedicated clusters – the job isn't writing code, it's making 10 really good decisions per day. without owned compute, weitzman says, even the best team hits a ceiling.

on top of that, they pay $100k extra per gpu to skip nvidia's delivery queue, getting rubin access a year before hyperscalers even list them – that's not a cost play, it's a time advantage.

source: @twentyminutevc podcast