"Just give me 1.5MW": the power bill for running 10,000 Opus-quality models at home
generativist · x · 2026-10-07
Responding to the recurring "local LLMs are good enough" debate, @generativist quips that local inference would only satisfy him if it delivered Opus 5.5-quality models running 5,000–10,000 in parallel — so, "conservatively," he'd settle for 1.5MW of residential power. A tongue-in-cheek way of quantifying the real compute gap between home setups and frontier cloud inference.
More from Fun
- Training two 'small language models': 9-month pretraining, reward hacking via snacks — Kyrannio · 2026-10-07
- Non-coder uses Claude to render himself in 23 code-drawn art styles — AlchainHust · 2026-10-07
- slopTV is coming: an all-AI music video channel rebuilt '90s style — max_paperclips · 2026-10-07
- Typewolf remains a rare gem for finding good typography in the age of AI slop — fernandorojo · 2026-10-07
- Dioscuri shares screenshot of AI assistant stirring controversy in the family agent group chat — dioscuri · 2026-10-07
- A witty sign-off riffs on Wittgenstein's ladder: this exchange ends with a walk — _onionesque · 2026-10-07