xAI brings Colossus 2 online — ~1-million-GPU supercluster now pre-training Grok 5

xAI says Colossus 2 is the world's first roughly one-million-GPU supercluster, with about 1.02 million NVIDIA GB300 GPUs split across sites in Memphis and Atlanta, drawing 1.4GW of total power and using 140kW-per-rack liquid cooling. Grok 5 is reported to be pre-training on the cluster now, with completion expected by August 2026.

The scale is a step-change in the compute arms race. Engineers reacting online called it 'the day the curve bent again,' and the move underscores Elon Musk's strategy of out-building rivals on raw infrastructure. Anthropic reportedly countered within hours with a $12B Amazon Trainium fleet expansion, signaling that the frontier-lab competition is increasingly a contest over who can stand up the most accelerators fastest.

The buildout also raises hard questions about power and siting: 1.4GW is comparable to a large nuclear plant's output, and xAI's rapid Memphis expansion has already drawn scrutiny over local grid and environmental impact. Whether Grok 5 delivers a corresponding capability jump — and whether xAI's distribution edge through Tesla and X can monetize it — remains the open question as the August training target approaches.