xAI Pivots to GPU Landlord, Renting Capacity to Anthropic and Google

xAI has signed massive GPU leasing deals with two direct competitors: $1.25bn/month to Anthropic for 300MW (~220k GPUs) at the Colossus 1 site in Memphis, and $920mn/month to Google for 110k GPUs. Both contracts allow 90-day cancellation after an initial lock-in. The Anthropic arrangement let that company roll back the peak-hour rate limits it imposed earlier this year to manage a severe compute crunch during overlapping European afternoon and US morning demand.

The economics are striking. Power, often assumed to be a major cost, runs roughly $90-160mn/year against ~$15bn/year in revenue from the Anthropic deal alone — about 1% of the top line. At that rate, xAI would recoup the entire capex within 18 months while still holding hundreds of megawatts of usable GPU capacity. SpaceX-style execution speed is the real edge here: Colossus 1 was built in 122 days, while hyperscaler and OpenAI Stargate projects remain years from completion, with the UAE buildout now also exposed to drone strikes from the Iran conflict.

The trade-off is Grok’s positioning. Renting the bulk of training and inference capacity to rivals signals xAI is quietly retreating from frontier-lab ambitions, likely because Grok inference demand fell short of projections. Layer on Google’s stake in SpaceX and the upcoming IPO, and there’s clearly some valuation engineering at play — but the underlying compute shortage and SpaceX’s buildout advantage are real. The result looks less like a frontier AI lab and more like a datacenter REIT with a model team attached.