Epoch: Huawei can't catch Nvidia this decade, GPT-6 Astra sets records, data center power doubles every 10 months
Epoch AI · rss · 2026-09-13
Epoch AI's brief covers: Huawei's most powerful chip is half an H100 and it will ship under 4% of Nvidia's AI compute in 2026, almost certainly unable to catch Nvidia this decade; US GDP statistics miss most of Nvidia's value, understating growth by 0.3 points; GPT-5.6 shows quadratic long-context latency scaling vs near-linear for Claude 5, hinting at architectural differences; data center power records double every 10 months (xAI Colossus 2 at 950 MW); GPT-6 Astra tops the ECI among 247 models, saturates FrontierMath Tier 4 at 98%, and solves 2 of 68 problems on the new FrontierMath Erdős benchmark; OpenAI grew compute 17-fold in two years.
More from Infra
- Team With 2x A100 Grant Open-Sources Roadmap for Distilling Offline Edge Models — Lerok-Persea · 2026-09-13
- Prediction: hyperscalers will tighten AI capex in 2027 — pdamodaran · 2026-09-13
- llama.cpp launches llama.app: frontier AI fully local with one command — ngxson · 2026-09-13
- Qwen on M2 Ultra: latest oMLX update brings substantial local inference speedup — Thrumpwart · 2026-09-13
- Running a 177B MoE Model on Dual 3090s and an Old EPYC: 38 tok/s for $800 — ludos1978 · 2026-09-13
- Google rewrites SERP links to opaque google.com/goto redirects, raising the cost of scraping results at scale — jedisct1 · 2026-09-13