A tweet argues closed models may cost $10,000 in compute for every megabyte of weights
JosephJacks_ · x · 2026-07-27
A tweet argues that Anthropic and OpenAI’s business model is absurdly expensive: using current SOTA methods, each megabyte of final closed-model weights may require roughly $10,000 in compute to train and serve.
The author’s back-of-the-envelope math claims a 10T-parameter model needs about 30 PB of uncompressed training data, which compresses into roughly 5 TB of model weights on disk—about a 6,000:1 compression ratio. That implies around $10 billion of compute per terabyte of ultimate model size, likened to building-sized 1940s mainframes spitting out enormous FLOPs.
More from Infra
- AI chip startups are attacking Nvidia’s data-moving bottleneck, with a $58B private-paper stack — FinanceYF5 · 2026-07-27
- AMD and South Korea Partner on Heterogeneous Computing and Local NPU R&D — JungWooHa2 · 2026-07-27
- NVIDIA and dozens of firms back a letter arguing open frontier models matter — ying11231 · 2026-07-27
- CXMT reportedly jumps nearly 500% on market debut amid AI memory-chip demand — Polymarket · 2026-07-27
- Local vs. Cloud: Evaluating image generation costs for indie game devs — Simple-Evidence-9125 · 2026-07-27
- Chart shows LLM context windows growing far faster than RAM or CPU clocks — nbaschez · 2026-07-27