ETH's Hoefler previews SC26 keynotes: weather-model scaling laws and AI infrastructure spend

thoefler · x · 2026-11-16

Torsten Hoefler (ETH Zurich/ADIA Lab) previews two SC26 keynotes: 'AI for Science' at the HPC4EES workshop and 'AI infrastructure' at ExaMPI'26. Slides show real findings: weather-model scaling-law comparisons across Aurora, GraphCast, SFNO, Pangu and AIFS (1–90 TB training data) — GraphCast and Aurora scale best, GraphCast needs fewer FLOPs but runs slower (arXiv:2602.22962) — plus WSJ data showing AI datacenter build-out at 3.63% of GDP, exceeding canals, railroads, electrification and telecom combined.

Original post →

More from Infra

Infra channel →