Gemini 4 Hydrogen tops Argon, Huang exits Nvidia

November 12 might go down as the most chaotic 24 hours in AI history, simultaneously upending frontier chip leadership, model-routing economics, and automated reasoning. When workhorse models beat frontier flagships on default settings and startups raise billion-dollar rounds to automate AI R&D itself, the bottleneck shifts decisively from human researchers to raw power and formal verification.

• –Gemini 4 Hydrogen replaces Google's Flash tier at $1.50/$7.50 per million tokens with a 1M context window and 2.5x faster execution than Gemini 3.8 Flash, scoring 51 vs Argon's 50 on the Artificial Analysis Intelligence Index at default high thinking (and scaling to 54 at xhigh) so developers can route everyday coding and subagent loops to the fast tier without losing capability
• –Jensen Huang's immediate resignation from Nvidia's CEO role and board — installing longtime CFO Colette Kress as interim CEO and Harvey C. Jones as independent chair amid probes into workplace and suspected intelligence-linked relationships — injects sudden governance and export-policy uncertainty into the AI hardware supply chain
• –Axiom Math's kernel-checked Lean 4 formalization of Faltings' theorem (built via Bombieri's proof atop Mathlib and the formalized Stacks Project with zero sorrys or custom axioms) shows AI-driven formal verification can conquer landmark arithmetic geometry, a capability directly transferable to high-assurance software engineering
• –Mirendil's $1B Series A at a $5B valuation, co-led by Kleiner Perkins and Clearlake Capital after a single researcher used autonomous agents to pretrain a competitive coding model from scratch, validates recursive agent-led ML research
• –SpaceX's $74B supercompute commitment — $24B for self-powered training clusters at Starbase, Texas, and $50B for a dry-cooled campus in Cochise County, Arizona — underscores that off-grid power generation and land scale now dictate frontier compute capacity