xAI's surprise deal with Anthropic on Wednesday, in which the Claude maker bought all of the compute capacity at xAI's Memphis-based Colossus 1 data centre, is being framed by Elon Musk as a tidy capacity-balancing trade. xAI moved Grok training to the larger Colossus 2 site, the argument runs, so Colossus 1 was spare. That framing leaves out the more interesting question. Why would xAI sell roughly 300 megawatts of state-of-the-art training capacity to one of its closest competitors at all?
The deal is reportedly worth billions of dollars and big enough that Anthropic was able to immediately raise usage limits across its product line on the strength of it. For Anthropic, the read is straightforward. The lab has been visibly capacity-constrained for months, Claude usage has been growing on a curve steep enough to keep buying more compute clearly worth the price, and a 300 MW slug of Hopper-and-Blackwell-class infrastructure that is already commissioned and running is a far better deal than waiting eighteen months for a greenfield build. Anthropic's behaviour is a textbook indication that the model is on a capability and demand curve that justifies the marginal compute.
The xAI side of the trade is harder to square with the "we are a frontier model lab" story Musk has told publicly for the last two years. Frontier labs running ahead on capability do not sell their training capacity to direct competitors. They hoard it. They reserve it. They use it to widen the gap. xAI did the opposite.
What reselling compute usually means
Compute-reselling deals at this scale are rarely what a frontier model lab does when its model business is going well. The cleanest recent precedent is Inflection AI in March 2024, which sold most of its team and its compute commitments to Microsoft after Pi failed to find product-market fit at consumer scale. The Inflection shell that remained pivoted to enterprise services. The pattern, broadly stated, is that a lab raises capital on a model thesis, accumulates compute to train the model, finds the model cannot compete with the better-capitalised frontier alternatives, and discovers that the data centres it bought have a separable second life as rental capacity. The compute pivot tends to come second, and it tends to come quietly.
xAI's version is louder, and it is happening at a higher valuation than most precedents, but the structural shape is the same. Grok usage has reportedly fallen sharply since the image-generation incidents earlier this year, the company's announced model releases have not narrowed the perceptible gap to Claude or GPT, and the public narrative has migrated steadily from "Grok will beat them all" toward "xAI's data-centre footprint is the asset that matters". The Anthropic deal is what the second narrative looks like in cash form.
The deal is the comparative-capability disclosure
Read the trade as a market exchange and it is more informative than any benchmark headline. Anthropic decided that buying Colossus 1's full output for years was a better use of billions of dollars than waiting on Anthropic's own pipeline of new sites. xAI decided that selling that same output to Anthropic was a better use of Colossus 1 than training Grok on it. Both decisions are individually rational. Together they are an unambiguous market signal that Claude is on a steeper capability and demand curve than Grok.
This is the inference Musk's framing of the deal works hardest to obscure. The "we have Colossus 2 now, we did not need Colossus 1" line is technically true and economically irrelevant. A lab confident that its next training run would produce a market-defining model would absolutely want both sites running on its own workloads. The fact that xAI was willing to monetise one of them at any price tells you what the company thinks the marginal compute would have produced if pointed at Grok.
Intermediated by Anthropic
There is a second, structural problem for xAI in this deal that is not about cash flow at all. The customer-facing AI product running on Colossus 1 from now on is Claude, not Grok. The chips xAI procured from Nvidia, the substations it negotiated with the Tennessee Valley Authority, the engineers it hired to bring the site online, are now infrastructure that powers Anthropic's growth. xAI has positioned itself, voluntarily, as a sub-tier supplier in someone else's stack.
This is exactly the position the better-resourced AI integrators have refused to occupy. Last month Sundar Pichai admitted on a Google earnings call that Google Cloud revenue was lower than it could have been because the company was capacity-constrained, and that when forced to choose between renting GPUs to external customers and using them for Google's own AI development, Google chose its own AI development. Meta, when it hit the same constraint, spun up an entirely new internal cloud apparatus to keep the chips inside the company. Both made the choice that xAI has just declined to make. Frontier integrators preserve compute for the product layer because the product layer compounds capability advantages. Neoclouds rent the compute and capture none of that compounding.
The pre-IPO rationale
There is also a corporate-finance reason the trade looks more rational once you look at xAI from the SpaceX vantage point rather than the pure model-lab vantage point. xAI is now part of SpaceX following the share-exchange combination Musk completed last year, and SpaceX is publicly reported to be moving toward an IPO. The Anthropic deal lands at exactly the moment that revenue line items matter most, and it has the shape of a transaction that, charitably, balances the books, and uncharitably, dresses up the prospectus.
Pre-IPO companies routinely accept revenue-shaped opportunities they would otherwise discount, and the public-market story SpaceX can now tell is materially different from the one it would have told without this deal. A multi-billion-dollar Anthropic compute contract, signed and partially recognised before the prospectus is filed, transforms xAI in the eyes of public-market investors from "frontier AI lab burning cash on a model that has not yet won" into "AI infrastructure operator with a multi-year contracted revenue tail from a top-tier frontier customer". For roadshow purposes, the difference is enormous, and importantly it does not require the underlying business to actually be the second thing.
The contract revenue is also unusually cheap to produce. Colossus 1 is already built, its substations are already energised, depreciation on the asset base will run regardless of who is using the chips, and converting unused capacity into Anthropic-paid revenue is effectively pure margin against a sunk cost. The accounting picture in the prospectus would, by design, look dramatically healthier than the underlying strategic picture. From a public-market investor's perspective, that is precisely the kind of one-off revenue uplift that should be discounted heavily before pricing the stock. From SpaceX's perspective, that is the point.
The valuation gap is the verdict
The market has already partly absorbed the implication. xAI was valued at $230 billion in its January funding round, on a story that is largely an integrated AI lab story. CoreWeave, the listed neocloud that operates a comparable amount of GPU capacity, is worth less than a third of that. The valuation gap is what the market pays for the difference between "we are a frontier model lab" and "we are a GPU rental business". Each step xAI takes toward the latter category narrows the gap from the wrong direction.
The remaining xAI ambitions read thinner once the compute pivot is read as the actual pivot rather than an opportunistic side trade. The Terafab in-house chip project will erode some of Nvidia's pricing power but will not change the basic neocloud-economics math. The orbital data centre project Musk discussed at the February all-hands sits at a 2035 timeline that cannot be a serious capacity strategy for the model business that needs to compete this year. The Cursor partnership positions xAI as a coding-tool compute supplier rather than a coding-tool builder. The Macrohard project, the digital-twin extension of computer-use research, is a long-horizon bet that needs committed compute precisely of the kind xAI just sold to Anthropic.
What to watch next
Three signals matter from here. First, whether Anthropic's near-term raised usage limits hold, or whether they are pulled back once the lab has worked through the immediate Colossus 1 backlog. The first is the bullish read of Claude's capability curve; the second would suggest the deal was as much about narrative as about compute. Second, whether xAI announces additional capacity commitments at Colossus 2, or whether it begins quietly leasing slices of that site as well. Third, whether the next Grok release narrows or widens the perceptible gap to the frontier. The deal's logic implies the gap is widening. A surprise capability release that contradicted that read would be the strongest argument for the "Colossus 1 was genuinely spare" framing.
The structural picture, for now, is the cleanest one: an AI lab that has decided, at least partially, that it would rather monetise its data centres than continue burning them to chase Claude and GPT. That decision is rational. It is also, historically, what failing model labs do. The neocloud category xAI is moving into exists because labs like Anthropic need somewhere to point demand. xAI just announced that it would prefer to be the somewhere.
