800 Blackwells vs 100,000: can labs without GPUs survive the open-weights race?
altryne · x · 2026-10-13
altryne highlights Peter Gostev's question: is there any hope for labs that can't get GPUs, as the compute gap widens to 800 Blackwells vs 100,000?
This week's big open-weight announcements: Reflection's 501B Beam, Mistral's 1T "Le Chonk", and Aleph Alpha's Kolibri — but two are still just announcements. "Drop the weights."
More from Infra
- TSMC September Revenue Jumps 54.6% YoY, Q3 Hits $46.7B on AI Demand — Beth_Kindig · 2026-10-13
- Open-source warmctl restores a warmed-up vLLM engine in ~37s on 8×H100 — kbesplv · 2026-10-13
- NVIDIA open-sources SMLIB, its NURBS-based CAD kernel, after 20-year partnership — Promptmethus · 2026-10-13
- STAPLE adds static sparse memory to pretrained LLMs with 50% less serving energy than MoE — BeidiChen · 2026-10-13
- OCP chatter: tighter scale-up/out compute domains could ease the race to ~500 GPUs per rack — BenBajarin · 2026-10-13
- NVIDIA releases NVFP4 quantized Qwen-Image with 12B-size weights on Hugging Face — TheZachMueller · 2026-10-13