Dev rants against GPU-stacking: DeepSeek's <1GB RAM per 1M context shows efficiency wins
gnukeith · x · 2026-11-22
Developer gnukeith criticizes the "just buy 80 more GPUs" approach, arguing it imposes no necessary constraints and benefits labs that centralize compute. He praises DeepSeek's opposite path — making base models smart but extremely efficient — highlighting that its 1M-token context needs under 1GB of RAM, a huge scaling win for self-hosting and neoclouds alike. He also mocks local-AI demos fitting big models on one GPU with only 256-token context windows.
Related event: Developer criticizes GPU-stacking approach, praises DeepSeek's efficiency(2 posts)→
More from Infra
- Lenovo's enterprise AI infra revenue jumps 200%, outpacing cloud giants, $80B pipeline — shashib · 2026-11-22
- Wiretok: open-source local monitor for Antigravity compaction headroom and API spend — phosphorhouse · 2026-11-22
- Running an unquantized 61GB BF16 30B model on a 12GB RTX 4070 SUPER at 0.8 tok/s — Ill-Organization-38 · 2026-11-22
- Musk details Tesla's in-car ~1B diffusion world model running at 36 fps in production — flngr · 2026-11-22
- Nvidia plunges 14% after Reuters reports SEC probe into its accounting practices — mark_k · 2026-11-22
- ING fails to sell $500M data center loan bundle, forced to sell loans one by one — churchkey · 2026-11-22