xAI

Contribution to AI

Compute stopped being something a laboratory bought and became something it built. The Memphis cluster called Colossus went from an empty Electrolux factory to 100,000 Nvidia H100s training on a single RDMA fabric in 122 days, then roughly doubled within months, with Tesla Megapacks installed to absorb the power swings that synchronised training imposes on a utility grid. Whatever one makes of the models trained on it, the schedule itself was the demonstration: site selection, substation capacity, liquid cooling and network fabric became frontier-lab competencies rather than someone else's procurement problem, and every competitor's build timeline was measured against it afterwards. The bill arrived publicly too — unpermitted methane turbines feeding the site drew Clean Air Act litigation from the NAACP and the Southern Environmental Law Center, which turned datacentre siting into a civil-rights dispute conducted in neighbourhood terms rather than a technical footnote. The March 2024 release of Grok-1 mattered separately: a 314-billion-parameter mixture-of-experts base checkpoint, weights and architecture under Apache 2.0, larger than any openly licensed model then available, released raw and unfinetuned. Wiring an assistant directly into a live social platform was the third experiment, and its publicised failures supplied the field with unusually concrete evidence about what a deliberately permissive content policy produces when the model sits on top of a firehose read by millions.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).