NVIDIA's $1B Single-Model Training Forecast Landed Years Ahead of Schedule

IgorCarron · x · 2026-09-05

Igor Carron resurfaces a 2021 Next Platform piece where NVIDIA's Bryan Catanzaro predicted a company could spend $1B in compute to train a single language model by around 2026. He notes the milestone arrived early: customers were already willing to spend $1B on LLMs in 2023 — even NVIDIA's own scaling predictions proved too conservative.

Original post →

More from Infra

Infra channel →