Anthropic Buys xAI's Colossus 1 Capacity — A Shift Toward a Neocloud Model

Anthropic has purchased all compute capacity at xAI’s Colossus 1 data center — roughly 300 megawatts — a deal that monetizes idle infrastructure immediately and reframes xAI as a potential provider of leased GPU capacity.

Anthropic has bought out all of the compute capacity at xAI’s Colossus 1 data center — roughly 300 megawatts — in a surprise agreement that immediately converts idle training capacity into revenue and lets Anthropic raise its usage limits. The arrangement monetizes Colossus 1 at once and has been described in reporting as large enough to be worth billions of dollars, turning a dormant asset into a commercial product for xAI.

Elon Musk said xAI had already shifted most of its model training to a newer facility, Colossus 2, removing the operational need for Colossus 1. That reallocation reflects xAI’s consumer — product focus — led by its Grok models — which, after a drop in Grok usage tied to earlier image — generation issues, no longer required both campuses at full capacity.

The deal is notable because it runs counter to how many major AI builders currently handle capacity. Companies often reserve GPU fleets for internal development: Google’s chief executive recently said Cloud revenue has been limited by being “capacity constrained,” and Meta has constructed a dedicated Meta Compute stack to secure GPU supply rather than rent it out. By contrast, xAI’s sale of an entire data center block is more characteristic of a compute provider selling wholesale access.

Viewed commercially, the move positions xAI more like a neocloud: an operator that buys GPUs at scale and leases compute to external model developers such as Anthropic. That business is capital — and supply — constrained, dependent on chip vendors and cyclical demand, and generally faces tougher economics than a pure software or consumer AI vendor. Leasing capacity trades predictability of product revenue for the capital intensity and margin pressures of infrastructure operations.

The swap also highlights how valuation and hardware strategy interact. xAI raised a $230 billion valuation in a January funding round, while specialist compute providers that focus on leasing capacity, such as CoreWeave, trade at a far lower multiple. Musk’s longer — term ambitions — including custom chips manufactured at a planned Terafab and even deploying some data centers in orbit by 2035 — would alter parts of the supply and cost equation but would not eliminate the fundamental capital and vendor dependencies of providing large — scale GPU capacity.

For AI builders, the near-term consequence is increased availability of large — scale GPU capacity outside the major public clouds and a paying commercial customer for xAI’s surplus infrastructure, which could accelerate model training for groups that buy that capacity. Strategically, the deal signals that xAI may prioritize infrastructure monetization to fund expensive ambitions (orbital centers, custom silicon and product development), even if that shifts emphasis away from owning every cycle for in-house model work.