xAI Is Becoming a Datacenter REIT, and Anthropic Is Paying the Rent

Anthropic is paying xAI roughly $1.25 billion every month to rent a single 300-megawatt block of GPUs inside a Memphis supercomputer that the company built in 122 days, and that one fact has quietly reorganized what xAI actually is.

The lease that rewires the story

On June 8, 2026, the independent analyst Martin Alderson published a post titled "xAI's New Rental Business" on martinalderson.com. The argument is compact and, once stated, hard to unsee. "xAI is now looking more like a datacentre REIT with a frontier lab attached, rather than the other way around," Alderson writes. A real estate investment trust, for readers who do not spend their afternoons reading 10-Ks, is a company that owns buildings and rents them out for predictable monthly cash. The shorthand matters because it is the lens through which Alderson is asking everyone to look at xAI's Memphis operation.

The piece is built around two leases. The first is the one that anchors the story. xAI is leasing a 300-megawatt block of its Colossus 1 facility in Memphis, Tennessee to Anthropic, the maker of the Claude family of models, for roughly $1.25 billion per month. That works out to about $15 billion per year flowing from Anthropic to xAI for compute capacity. Inside that 300-megawatt block sit approximately 220,000 Nvidia GPUs drawn from the H100 and H200 generations. The second lease, signed with Google, runs at about $920 million per month for roughly 110,000 GPUs, widely understood to be the newer GB200 silicon. That is why the per-GPU rate on the Google block is much higher than the Anthropic block: about $8,400 per GPU per month for the GB200 capacity, against roughly $5,700 per GPU per month for the Memphis H100 and H200 tenancy.

There is a third character in this story that is easy to forget. In February 2026, the xAI–SpaceX merger was completed, folding xAI into Elon Musk's rocket company ahead of the long-expected SpaceX public offering. The cash that Anthropic and Google are now sending Memphis lands, in effect, on SpaceX's pre-IPO ledger.

Colossus 1, by the numbers

Colossus 1 is the Memphis supercomputer site that hosts these GPUs, and it is the operational fact that makes the rest of the story possible. xAI stood it up in 122 days, a build window Alderson describes as an extraordinary infrastructure-execution feat that hyperscalers like Amazon Web Services or Google Cloud could not match. The total capital expenditure to build the GPU clusters, supporting power, cooling and the surrounding site is estimated at around $40 billion.

The unit economics are what give Alderson's REIT framing its weight. At combined Anthropic and Google lease revenue of more than $2 billion per month, xAI recovers that roughly $40 billion build cost in about 18 months, before counting any inference revenue from its own Grok product. The payback is not promised, and the standard cancellation term in this class of multi-billion-dollar compute lease is a 90-day notice period after an initial lock-in, which gives tenants like Anthropic an exit if model demand falls. But the slope of the curve is the point. A wholesale compute landlord that can sign two tenants at this size, in the months immediately after pouring the concrete, is operating closer to a real estate logic than a research logic.

Power, the input that usually wrecks these spreadsheets, is the surprising part. Tennessee grid electricity at Memphis costs roughly 6 cents per kilowatt-hour, and the facility also runs on-site natural-gas turbines that lower the marginal cost of power further. The estimated annual power bill for the 300-megawatt Anthropic block sits between about $90 million if the on-site gas turbines carry the load and $160 million on grid power. Against the $15 billion the block generates in lease revenue, electricity is roughly one percent of revenue. That ratio is closer to a colocation business than a frontier laboratory burning capital to train.

Why Anthropic chose to rent

The natural question is why Anthropic, a company with its own cloud relationships and its own compute strategy, is wiring more than a billion dollars a month to a direct rival. Alderson's reading is that capacity, not preference, drove the decision. Earlier in 2026, Anthropic publicly imposed peak-hour usage restrictions on Claude Code subscriptions, with usage between 5 a.m. and 11 a.m. Pacific Time, or 1 p.m. to 7 p.m. Greenwich Mean Time, counting more heavily against user limits. Alderson links those peak-hour caps to capacity strain, what he describes as "serious capacity problems, especially early afternoon onwards in Europe," which the new Memphis lease is meant to relieve.

The competitive backdrop is unforgiving. "At some point you end up having to ration users further, which definitely is far from ideal when you have both Google and OpenAI breathing down your neck for customers," Alderson writes. Rationing Claude is not a long-term strategy when rivals are not rationing theirs, and building a Memphis-class site from scratch is a multi-year exercise for almost any operator who is not xAI or SpaceX. Renting is, in effect, the only door that opens this calendar year.

Siting matters for reasons that have nothing to do with electricity prices. Earlier in 2026, Iranian drones struck datacenters tied to the Stargate buildout in the United Arab Emirates, an incident that hardened the case for siting frontier compute in lower-risk locations such as the U.S. Midsouth. The Memphis campus is, among other things, a long way from a contested airspace.

A landlord whose biggest tenants are its rivals

Step back from the spreadsheet and the strategic picture becomes more uncomfortable. The largest single-site GPU supercomputer in the world is now a wholesale compute landlord whose biggest tenants are direct competitors of its own Grok product. Alderson argues that Grok has effectively gone into "a serious retreat" from being a frontier-class lab, because GPU capacity that was nominally pencilled in for Grok training is now being rented out to direct competitors. The 220,000-GPU block in Memphis is not training the next Grok. It is serving Claude.

Alderson is careful not to call this a defeat. His framing is that three things are simultaneously true. SpaceX's pre-IPO financials benefit from the recurring lease revenue. The broader market has a genuine GPU shortage that makes Memphis capacity unusually valuable. And xAI has a real, durable advantage in building and energizing large datacenters quickly, an advantage that survives whatever happens at the model layer. The question the piece leaves open is whether that triangle adds up to a frontier AI lab or to a landlord with a chatbot attached.

It is worth saying plainly what the deal does not do. It does not put Anthropic out of business; the company is paying for capacity, not surrendering ground. It does not foreclose Grok's research roadmap; xAI can build Colossus 2. It does not guarantee the 18-month payback, because lease cancellations and demand shocks are real. What it does do is reset the prior. A company that was supposed to be measured on benchmark scores is now measurable on megawatts, monthly rent and the cost of Tennessee electricity.

Sources

• Martin Alderson, "xAI's New Rental Business" (June 8, 2026)