COMPUTE REAL ESTATE

Epoch AI counts about 30 GW of frontier AI compute today, doubling roughly every seven months. US electricity generation is growing by around 120 TWh a year (2.8% in 2025) — healthy by historical standards, but a fraction of what that doubling implies. The result is a queue: projects wait not for GPUs or money but for the utility to say yes. Any site with a signed interconnection agreement is worth more than the building on it.

Operators have stopped waiting for the grid. Amazon has contracted ~2.9 GW of nuclear output (1.92 GW from Talen’s Susquehanna plant and 960 MW of X-energy small modular reactors). Meta’s 1 GW-plus Prometheus campus in Ohio is slated to come online in 2026 with on-site gas. Battery storage is being added at 20–25 GW a year globally (IEA) to smooth the load. In each case the operator is behaving like a landlord securing utilities for a tenant, not like a software company.

Data Center Watch tracked $162B across 36 projects blocked or delayed by local opposition by mid-2025; by Q1 2026 the tally was ~$130B across 75 projects in a single quarter. Water use, noise and electricity bills are the objections. Jurisdictions with land, water and spare generation — and public support for using them — become the scarce resource.

Every constraint on this page is an advantage for the region. Cheap, abundant generation (Saudi Arabia’s record solar tariff is ~$0.0104/kWh), land without neighbours to object, and permitting under a single authority. HUMAIN’s targeted 6.6 GW would be, on its own, more than a fifth of today’s global frontier fleet. The question is no longer whether the Gulf can host compute, but on what terms it leases it.

Mercatus AI’s model (May 2026) frames the economics the way a landlord would: the building lasts 15–20 years, the chips inside it three to six, and the lease that pays for both typically runs two to seven years. The operator’s return therefore depends on three things — utilisation, the contracted rate, and whether the tenant renews before the chips are obsolete. The bulk of the capital is debt-financed against that contract, so a strong tenant lowers the cost of every gigawatt.

Only about $8.5B of the $38B gigawatt is bricks, power and cooling. That inversion — the fit-out worth four times the building — is what makes this asset class new. It rewards owners who can refill the building with the next generation of chips, and it punishes anyone who financed the chips over the life of the building. Every structure in Section 05 is an attempt to match the money to the right lifespan.

Who: CoreWeave, Nebius, Crusoe. How: the operator controls the sites (owned or on long leases), buys the chips, raises the debt and carries the depreciation. Reward: keeps the entire margin and the terminal value of the land and power. Cost: balance sheets that now rival utilities — CoreWeave carries ~$35B of debt — and full exposure if a tenant does not renew. Capital spending guidance for the two listed leaders alone is $55–64B for 2026.

Who: Fluidstack and a growing set of operators built on the same idea. How: a landlord — often a former bitcoin miner with powered land — owns the site and builds the hall; a hyperscaler or the tenant backstops the landlord’s lease; the operator designs, fills and runs the cluster and holds the customer relationship. Reward: hyperscale capacity without hyperscale debt. Cost: the operator shares the economics with the landlord and the guarantor.

Contracted backlog is the number that matters, and the two listed leaders alone hold ~$144B of it. Backlog is the lease roll of this asset class: it tells you how much rent is already promised before a single new building is financed. Private operators are now pricing off the same metric, which is why the public scorecard in Section 07 sets the valuation benchmark for the whole group.

Every operator on this list depends on a handful of tenants, and several on one. That is normal for a new property market — the first office towers in any district are anchored by a single large lessee — but it means a tenant’s decision to build its own capacity, or to renegotiate, moves an operator’s value more than any technology change. Section 11 returns to this.

Every name on this page is profitable before depreciation and loss-making after it. That is not an accounting trick; it is the asset class. A landlord who has just built a tower shows the same pattern in year one. The question investors are really asking is whether the chips will still be earning rent when the depreciation stops — which is a question about contract length and renewal, not about margins.

The market pays for contracted backlog and punishes concentration. Nebius, with balanced debt and cash, has more than doubled; CoreWeave trades at roughly twice its IPO price despite $35B of debt because $104B of rent is signed; IREN and Applied Digital, still carrying mining legacies and single-tenant exposure, have gone sideways. Compute real estate is being priced like real estate: on the quality of the lease.

Michael Burry’s critique — that the largest operators are understating depreciation by an estimated $176B over 2026–2028 by extending chip lives — is the correct question to ask of any leveraged property market: is the rent roll longer than the asset life? OpenAI’s decision to scale its stated commitments from ~$1.4T to ~$600B is the other side of the same coin. Neither changes the direction of the market; both change the price at which it should be financed.

Gulf institutions are natural lenders and guarantors in this structure. They already own the power; providing the credit umbrella — lease guarantees, investment-grade facilities, anchor commitments — converts that advantage into a seat at the top of the capital stack, with real assets as security. The financing layer, not the chip, is where sovereign capital is best placed.

Starcloud is building data centres in space, where solar power is continuous and cooling is free by radiation. It has raised $450M to date and has flown its first hardware. The thesis is the purest form of this issue’s argument: if the scarce input is powered land, the answer may be to remove the land. It is early, and the launch-cost curve decides the timing — but it is the only venture that escapes the grid queue entirely.

HUMAIN, backed by Saudi Arabia’s Public Investment Fund, targets 6.6 GW of capacity behind ~$77B of planned investment, with ~$23B in its first phase; its Riyadh and Dammam campuses are targeted for 2026, with commissioning not yet publicly confirmed at the snapshot date. Abu Dhabi’s Stargate UAE is building a 1 GW campus with OpenAI and partners. Both sit on the cheapest utility-scale power in the world — the Kingdom’s record solar tariff is ~$0.0104/kWh.

SF Compute ($40M raised) runs an exchange where idle clusters are sold by the hour, setting the ~$1–2 spot floor referenced in Section 04. Prime Intellect ($130M) goes further, pooling distributed machines to train models across the internet. Neither owns a building; both make the asset class more liquid — the equivalent of short-let platforms for property.

Together AI ($8.3B valuation) and Groq (a $1.5B Saudi commitment, announced February 2025) sell model output rather than raw GPU time. They are the tenants of compute real estate as much as its operators — and their growth is the demand signal that keeps the rent roll full. Issue 02 covered why inference volume is compounding faster than prices fall.

Publicly named customers include Anthropic — which named Fluidstack as a partner in its US data-centre programme — Mistral, Meta and a set of leading open-model labs. The customer list is the clearest evidence for the model: labs that could rent from anyone choose an operator with no balance sheet of its own, because what they are buying is speed and operating competence, not real estate.

In March 2026 the company stepped back from a large European campus project it had announced a year earlier, choosing to concentrate on North American sites where power and guarantors were in place. For an asset-light operator that is the right reflex: it should never be the party carrying a gigawatt of unfunded land.

Sovereign and semi-sovereign entities own the two scarcest inputs — powered land and permitting authority. In the Fluidstack structure the landlord earns a long-dated, guaranteed lease and keeps the terminal value of the site. That is a real-estate return on an energy advantage, with an operator handling everything inside the building.

The credit umbrella needs an investment-grade name willing to backstop leases in exchange for priority capacity or equity-like upside. Gulf institutions are among the few balance sheets in the world that can play that role at gigawatt scale — and it places them at the top of the capital stack with real assets as security.

A six-month prefabricated build on a site with existing power is the fastest path from decision to revenue in the industry. For national programmes that have announced capacity targets on public timelines, an operator that can fill a hall in months — rather than a developer that will still be pouring concrete in two years — is the difference between a headline and a rent roll.

Because Fluidstack does not carry the hardware on its own balance sheet, it is less exposed to the depreciation question raised in Section 08. When the next generation of chips arrives, the operator refills the building; the landlord’s lease continues; the guarantor’s capacity is upgraded. The party that owns the ground is the party that benefits from every refresh.