Mistral hits 72% of the intelligence with just 3.8% of the GPUs, fan math claims

JFPuget · x · 2026-10-08

A viral comparison tallies "intelligence per GPU": Mistral 4 reportedly trained on 3,800 Grace Blackwell chips, vs 100,000+ for GPT-6 Astra, 200,000 estimated for Grok 4.7, and 500,000 estimated Trainium2 for Claude Opus 5.5 — concluding Mistral delivers 72% of the intelligence with 3.8% of the GPUs. Note most rival figures are estimates and the framing is playful rather than a rigorous benchmark.

Related event: Mistral Achieves '72% of Intelligence' With Just 3.8% of the GPUs, Sparking Debate(2 posts)→

Original post →

More from Fun

• Hackers Turn E-Readers into Always-On Muse Displays With MagSafe — FinanceYF5 · 2026-10-08
• "That's called exporting aluminum": AI circle jokes China's power glut is already being shipped abroad — teortaxesTex · 2026-10-08
• Bun creator regrets not reserving the bun lint and bun format subcommands — jarredsumner · 2026-10-08
• AI reading bedtime stories to kids sparks backlash: automating what 'should never scale' — CurieuxExplorer · 2026-10-08
• Roblox founder: GOTY used to take studios, now a prompt and a weekend — CurieuxExplorer · 2026-10-08
• "Who's ready for the Goodhart Singularity?" — a one-line jab at metric-gaming AI — burny_tech · 2026-10-08

Fun channel →