Energy and the power grid, not chips, are the real bottleneck for AI at 400k-GPU scale
kimmonismus · x · 2026-09-07
Responding to a discussion about large-scale compute deployments, kimmonismus argued that the paper math looks wild but "keeping 400k chips fed without the network falling over is the real nightmare," adding that energy and the power grid are the biggest bottleneck for AI scaling — not chips.
More from Infra
- Tracking one tennis ball with GPT-6 burns 7.87M tokens — Scobleizer · 2026-09-07
- Huawei's Kirin 9050Pro uses logic folding to cut NPU power 66%, run 30B MoE on-device — APPSO · 2026-09-07
- Signing TLS handshakes inside a TPM to protect machine identity — jedisct1 · 2026-09-07
- GLM 5.3 and Qwen 3.8 now run really well locally on single desktops — jasonkneen · 2026-09-07
- Same RTX 5080, 70% slower: renting GPUs across regions can silently cost you speed — Grouchy_Television75 · 2026-09-07
- Adding an RTX 2000 Ada 16GB (75W) to a gaming PC for local LLM inference under a 700W PSU — tableball35 · 2026-09-07