OpenAI could 7x its training compute tomorrow: why open-source models still trail by one generation
soumitrashukla9 · x · 2026-09-11
Peter Gostev argues open models feel close to proprietary ones only because leading labs are still using a tiny fraction of their compute for flagship training runs.
- Jensen Huang said Astra was trained on 100,000 Blackwells — roughly 7% of OpenAI's total compute. Using Epoch estimates, OpenAI likely has 3.4M H100-equivalents by HY 2026 (100k Blackwells ≈ 250k H100-equivalents at a 2.5x conversion).
- That means OpenAI (and presumably Anthropic) could 7x its next training run tomorrow and still keep 50% of compute for inference.
- Meanwhile labs like Moonshot or Alibaba can access perhaps 30–100k H100-equivalents (domestically or rented in Malaysia/the Middle East) — right in the class of models OpenAI/Anthropic train today, explaining why open weights keep catching up.
More from Infra
- OpenAI Agents API hits public beta; Cloudflare ships sandbox integration for cloud Codex agents — ritakozlov · 2026-09-11
- DeepSeek-V4.1-Flash lands on Fireworks: 552B MoE for coding and agents at 1/40th claimed cost — lqiao · 2026-09-11
- OpenAI Agents API Meets Cloudflare: Deploy a Logged Agent in 4 Minutes — craigsdennis · 2026-09-11
- Training a 6-Expert MoE GPT-2 From Scratch on a Single RTX 3090 in 8 Days — rasbt · 2026-09-11
- B3IQ Sells Eight Figures of GPUs in Two Weeks, Bets AI Infra Is a $100B Market — templecrash · 2026-09-11
- It Cost $100 in API Credits for an AI Agent to Install Free Software — MartinGTobias · 2026-09-11