Musk Concedes Anthropic Leads Grok, Then Orders 660,000 GPUs Before Year-End

A concession with a purchase order stapled to it

On September 26 Elon Musk conceded a race he has spent three years insisting he would win: Grok trails Anthropic's Claude models, Anthropic is the leader with no close rival, and he had misjudged that. His stated reason was arithmetic, not silicon. xAI is three years old against Anthropic's six ¹. One day earlier he had posted the remedy anyway: 660,000 more Nvidia GB300s at Colossus 2, in three waves of 220,000, taking the Memphis fleet from 780,000 GPUs to 1.44 million by late December↗ ².

The bill lands on public shareholders now

This order no longer rides venture money. SpaceX bought xAI in February, then went public on June 12 at $135 a share in the largest IPO in history↗ ³⁴. The listing's numbers show the machine being fed: $18.7 billion of 2025 revenue, a $2.6 billion operating loss, and more than $30 billion of capital spending across 2025 and early 2026 ⁵. The Blackwell waves are the next installment.

The yardstick is also a tenant

The rival Musk must catch already pays him rent. Anthropic signed for 300 megawatts of Colossus 1 at $1.25 billion a month↗ through May 2029, over $40 billion in all, which SpaceX's own S-1 describes as monetizing unused compute ⁶. Google adds $920 million a month from October and Reflection $150 million, more than $76 billion contracted through 2029, every deal breakable on 90 days' notice, and Musk has said he might need the capacity back ⁵. Anthropic and Google as paying tenants of Colossus is now part of the Memphis business model ⁸.

The turbines come out on a clock

Doubling a fleet means roughly doubling the power behind it. Colossus runs on 69 gas turbines that the NAACP and the Southern Environmental Law Center sued over as unpermitted↗; SpaceX says they come out by July 2027 as a permanent 1.2 gigawatt plant arrives, and its IPO filing budgets $2.8 billion for turbines over three years ³. The December wave, in Musk's words, needs luck ²⁷.

December 31 is the exam

Every promise now carries a checkable date: 220,000 GPUs operational this week, a Fable/GPT-6-level model within three months, pole position in six ². Grok 4.5's 53 percent on a coding benchmark is a team-years problem ¹. The wager underneath 1.44 million GPUs is that scale can amortize time.