Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

Compute Landlords: When AI Builders Become Rentiers

The companies that spent hundreds of billions building AI infrastructure just discovered a second business: charging everyone else to use it.

The Week the Builders Became Landlords

On July 1, 2026, Bloomberg reported that Meta is building a cloud business to sell its spare AI computing power. The unit, reportedly called Meta Compute, would offer raw GPU cycles and hosted access to Meta’s models, in direct competition with Amazon Web Services, Microsoft Azure, and Google Cloud. Wall Street’s reaction says everything about how investors read the AI buildout in mid-2026: Meta’s stock, down double digits for the year, jumped about 10% in a single day.

Think about what that price move means. The market didn’t reward Meta for a breakthrough model, a new product, or a research result. It rewarded Meta for finding someone else to pay for its chips. That instinct, chase whoever still needs to pay rather than whoever built the smartest thing, is the value-migration pattern in miniature: once the model goes cheap, the money collects wherever scarcity survives, and right now that’s the warehouse.

Meta guided $115 to $135 billion in capital spending for 2026, nearly double the previous year. At that scale, a GPU sitting idle between training runs is the most expensive paperweight ever manufactured. Nvidia’s accelerators depreciate like fish, not like real estate: each new generation (Hopper, then Blackwell, now Vera Rubin) cuts the economic value of the last one. If your chips aren’t computing, they’re rotting. Renting out the idle hours converts a depreciation problem into an income stream.

So the company that spent a fortune hoarding compute now wants to be your compute cluster’s landlord. And Meta isn’t even first.

Two weeks later the first prospective tenant surfaced, and it was a rival. On July 17 the New York Times and Reuters reported that Meta was in early talks to lease compute to Anthropic: up to $10 billion over two years, paid in monthly instalments, proposed by Anthropic in June, and cancellable by either side before the term ends. Meta declined to comment, and as of mid-September nothing has been signed, but the shape is already familiar from the deals below: the customer for spare frontier capacity is another frontier lab. The stranger detail is that the would-be landlord is itself one of the largest tenants on Earth. In April Meta expanded its CoreWeave contract to about $21 billion, running through December 2032, on top of a $12 billion five-year deal with Nebius signed in March for deliveries from early 2027, with an option on $15 billion more. A company renting some $48 billion of other people’s GPUs while marketing its own spare ones is not describing a surplus. It is describing a queue.

The Precedent: Renting Musk’s Fortress

The template came from Elon Musk. xAI built Colossus in Memphis into the largest AI training site on Earth, then got folded into SpaceX, which went public on June 12, 2026 at $1.77 trillion, the largest IPO on record. Buried in the S-1 was the detail that reframed the whole industry: the biggest users of Colossus are xAI’s competitors.

• Anthropic leases the entire output of Colossus 1, about 300 megawatts of compute, for $1.25 billion a month through May 2029. That’s a direct rival of xAI paying roughly $15 billion a year to train on Musk’s hardware.
• Google pays SpaceX $920 million a month from October 2026 through June 2029 for roughly 110,000 GPUs. A Google Cloud spokesperson called it “bridge capacity” because demand for Gemini outran even Google’s ability to build.

Combined, that’s about $26 billion a year in rent, flowing from two of the best-funded AI companies in history to a company they compete with daily. Read that sentence again. Google, which operates some of the largest data centers ever constructed, cannot build fast enough and is paying a competitor for the difference.

This is the detail that should recalibrate your intuitions: the tenants are not scrappy startups getting democratized access to compute. They are titans signing multiyear exclusives because the alternative is falling behind. When a giant rents your warehouse for three years, that’s not a spot market emerging. That’s scarcity being allocated by whoever owns the building.

Monetizing a Shortage, Not a Glut

“Spare capacity” suggests surplus, and surplus suggests falling prices. The market says otherwise. Rental rates for top-tier chips have been climbing again through 2026; H100 rates rose roughly 40% off their October 2025 lows as inference demand (much of it from AI coding agents) soaked up supply. Silicon Data’s daily H100 index, the closest thing compute has to a posted price, printed about $2.53 per GPU-hour in mid-September for a chip that is three generations old, and its third-quarter regional survey found hyperscaler rates down 26 to 54% since early 2025 while the independents’ rates recovered, so the two tiers are converging on a price rather than the price collapsing. Meta isn’t dumping excess inventory into a soft market. It’s selling scarce capacity into a shortage, at shortage prices.

That distinction decides who captures the value. In a glut, renters win: prices crash toward marginal cost, and compute behaves like a commodity. In a shortage, landlords win: prices reflect desperation, and the terms (duration, exclusivity, priority during contention) belong to the owner. Every signal in mid-2026, from the leases above to Morgan Stanley’s estimate that just four firms will spend about $630 billion on data centers and chips this year, points to the shortage persisting. The four biggest spenders are now positioning to own the ground everyone else rents.

The optimistic gloss is that compute is finally being “priced like electricity,” a metered input any business can buy. There’s something real here: a market price for compute is more legible, more contestable, and more efficient than compute locked inside vertically integrated silos. But finish the analogy. Electricity became a universal input because the meter’s owner got regulated. Your electric utility is a monopoly bound by rate cases, universal service obligations, and public oversight precisely because society learned what unregulated infrastructure owners do with essential inputs. Meta Compute has a meter. It does not have a rate case.

We’ve Run This Experiment Before

The 19th century built railroads; the railroads then discovered that owning the track was better business than running the trains. Farmers paid whatever the line to market charged, until common-carrier law forced rail into regulated, nondiscriminatory pricing. Standard Oil’s genius wasn’t refining; it was controlling the pipelines and rail rebates, taxing rivals’ product on the way to market. Electrification produced the same fight, resolved by utility commissions and rural cooperatives.

The pattern repeats because the economics repeat. Infrastructure with high fixed costs and concentrated ownership drifts toward rent extraction, and it stays there until some force (regulation, antitrust, cooperative ownership, or a technological end-run) pushes it into commodity pricing. The push has never come from the landlords’ goodwill. It took decades each time, and the Iron Law of Oligarchy explains why: whoever administers a scarce resource acquires interests of their own.

Compute is running the same script at fast-forward. The question the book keeps asking applies here with unusual precision: an input becomes genuinely abundant only when its owners are forced into utility-style commodity pricing rather than gatekept rents. Intelligence-per-dollar keeps improving, but access to intelligence is mediated by a shrinking number of counters, and the concentration of essential infrastructure in private hands is exactly how technological plenty curdles into political dependence.

The Fork: Utility or Toll Road

So does the landlord pivot bend toward unscarcity or away from it? Watch three levers, because they, not the press releases, decide the outcome.

1. Pricing structure. Commodity utilities converge on transparent, posted, nondiscriminatory prices. Toll roads run on negotiated, secret, discriminatory deals. There is real movement on the first kind: daily rental indices for the H100, H200, A100 and B200 now carry Bloomberg tickers, and exchange-traded GPU futures are being designed around them, which is what price discovery looks like when an input starts to behave like a commodity. But the Anthropic and Google leases are the second kind: bespoke, exclusive, and sized so only a handful of firms on Earth could sign them. If Meta Compute publishes a price sheet a mid-size lab can act on, that’s a utility signal. If its best capacity moves through private multiyear exclusives, that’s a toll road.

2. Interconnect and switching costs. Electricity is fungible; you can’t tell whose generator made your kilowatt-hour. Compute isn’t there yet. Proprietary interconnects, custom software stacks, and data gravity make switching landlords expensive, and every point of switching cost is a point of pricing power. Standards that make workloads portable across clouds do for compute what alternating current did for electricity. The landlords know this, which is why they’ll resist.

3. Who can be a tenant. The Labor Cliff makes this more than a business question. If intelligence is becoming the primary factor of production, then access to compute is access to economic participation, which is the argument for Universal Basic Compute: a guaranteed baseline allocation, provisioned the way the Foundation provisions housing or food. A rental market whose minimum viable tenant is a trillion-dollar corporation is the opposite of that. Watch whether sovereign compute programs, research allocations, and cooperative pools get real capacity, or whether the landlords’ book stays exclusively blue-chip.

There’s a fourth thing to watch, and it’s the cynical one: who pays the input costs. The rent flows up to the landlords while the electricity buildout that powers the buildings lands partly on household utility bills. Privatized rent, socialized grid costs. If that arrangement persists, the rentier structure isn’t an accident; it’s a policy choice.

The Unscarcity Read

The book’s three scenarios map cleanly onto this fork. In the Star Wars trajectory, compute stays a gatekept rent: four landlords, negotiated access, and a technological aristocracy defined by who holds leases. The Trojan Horse trajectory runs through the utility path: posted prices, portable workloads, public and cooperative capacity, and eventually compute allocations as civic infrastructure rather than corporate favor.

None of this requires the landlords to be villains. Railroads, oil trusts, and electric monopolies were run by people who considered themselves builders, and mostly were. The system drifts toward rent because rent is what infrastructure ownership pays, and it takes deliberate counter-design (Axiom IV: power must decay) to keep any of it pointed at abundance.

Meta Compute might genuinely push prices down; more supply usually does. But watch the levers, not the vibes. When the same four firms own the chips, the buildings, the interconnects, and the pricing, “compute priced like electricity” is only good news if someone remembers what we had to do to the electric company.

The factories of intelligence are built. The rent is being set now. Who gets to be a tenant is the whole game.

Related Articles

• Compute Clusters - The GPU arms race that built the buildings now being rented
• Universal Basic Compute - The case for compute as a guaranteed allocation, not a corporate favor
• Who Pays for AI’s Electricity? - The grid-cost side of the same fight: privatized rent, socialized bills
• Infrastructure Libertarianism - What happens when essential infrastructure answers to no one
• The Iron Law of Oligarchy - Why resource administrators always develop interests of their own
• Three Scenarios Analysis - Star Wars, Trojan Horse, or Patchwork World
• The Electron Gap - The energy constraint underneath the compute constraint
• The Commoditization of Intelligence - When the models themselves go cheap, the rent on the warehouse is where the margin hides
• Gateway to Intelligence: Who Owns the AI On-Ramp - The same rentier fight pointed downstream at the consumer: whoever owns the phone, the OS, and the wake word owns the on-ramp to every model

The Unscarcity blueprint argues that abundance is an engineering and governance problem: the technology is arriving, and the question is who controls it. Read the book or start with the preamble.