SpaceX’s newly combined AI business has begun leasing some of the world’s most concentrated Nvidia capacity to the same laboratories competing with its xAI arm — and the contracts suggest that scarcity, rather than simply scale, is becoming its most valuable product. Since SpaceX merged with xAI in February 2026, the company has turned the Memphis Colossus I and II data centres into an infrastructure business as well as an internal engine for AI development. Anthropic agreed to pay $1.25bn a month through May 2029 for about 325,000 Nvidia GPUs, according to SpaceX’s S-1 filing.

Google then agreed to pay $920m a month for roughly 110,000 GPUs from October 2026 through June 2029, filings and reporting said. Those commitments make SpaceX an unusually consequential new landlord in the AI compute market. They also expose a less comfortable fact for customers: the owner of the racks is no neutral utility.

It is building its own frontier-model business, while renting capacity to rival model makers.

A premium for flexibility Brett Winton, an analyst at ARK Invest, argued in an X thread on October 9 that the combined company’s advantage rests on its ability to accept shorter contracts than conventional “neocloud” providers. Such operators often need medium-term customer commitments to support the debt and capital spending required to build GPU campuses. SpaceX, Winton argued, has a stronger balance sheet, a relationship with Nvidia and, crucially, an internal buyer for unused capacity: xAI. That changes the economics of an empty server hall. A neocloud with no tenant still has financing, power and equipment costs. SpaceX can redirect spare chips to its own model training and inferencing, reducing the immediate pressure to lock customers into long agreements at lower prices. The thesis depends on continued expansion of its GPU fleet, rather than on a guaranteed margin.

“Revenue growth is insane. Never seen anything like it.” — Elon Musk, replying to Winton on October 10 Musk’s remark was not accompanied by independent confirmation of revenue beyond the company’s disclosed contracts and filings. But the disclosed deal values show why investors are watching: the two agreements alone amount to more than $2bn a month at their stated rates once both are fully active, subject to ramp provisions and cancellation terms.

ARK’s chart-based model, not company guidance, puts the proposed advantage in striking terms. It estimates a $50.5bn per gigawatt-year weighted average realised price for SpaceX between 2026 and 2036, against $26.2bn for a neocloud — a 93% premium, or about $24.3bn per gigawatt-year. ARK’s conversions put the Google agreement at $53.6bn per gigawatt-year and Anthropic’s at $33.1bn.

Winton has suggested blended infrastructure-as-a-service monetisation of $30bn to $40bn per gigawatt for much of the decade, potentially exceeding $50bn later.

Rich contracts, weak handcuffs

The same flexibility that may support premium pricing also limits the certainty of the revenue. Anthropic’s agreement can be terminated by either party with 90 days’ notice, reporting on the filing said. Google’s deal similarly allows either side to terminate with 90 days’ notice beginning in 2027.

The contracts are enormous, but they are not the kind of long, immovable commitments on which many data-centre builders prefer to base multibillion-dollar construction plans. Customers also have every reason to preserve alternatives. Anthropic and Google are not merely buyers of compute; they are major AI competitors, with their own infrastructure strategies and commercial stakes in not becoming overly dependent on Musk’s company.

The Information reported on October 1 that SpaceX’s AI unit had held exploratory talks over the summer with Microsoft about leasing capacity. Those were reported discussions, not a confirmed deal or term sheet. There is a further constraint in the equipment itself.

The value of a GPU cluster decays as newer chips arrive, even when demand remains fierce. A premium calculated per gigawatt can therefore be misleading if it treats electrical capacity as interchangeable across generations of hardware. A megawatt of older accelerators may not command the same price, or support the same workloads, as a megawatt of the newest Nvidia systems.

The scale question Even SpaceX’s physical scale remains contested. Data Center Dynamics has

reported about 340MW of actual capacity at the Colossus sites, against a 1GW nameplate figure cited elsewhere. That gap matters because ARK’s most dramatic comparisons are expressed per gigawatt-year. The company’s commercial promise depends not only on winning contracts, but on bringing usable power, cooling and current-generation chips online quickly enough to serve them.

“SpaceXsi,” the label used by Winton, is analyst shorthand for the combined SpaceX-xAI enterprise; it is not the company’s formal name. SpaceX brands its AI arm SpaceXAI. Anthropic’s stated interest in multi-gigawatt orbital compute, meanwhile, remains an expression of interest rather than a signed capacity contract.

The emerging model is still distinct.