AI Rents in Megawatts

In the second week of July 2026, one of the largest open marketplaces for rented computation carried the flagship servers of NVIDIA’s two most recent generations. The newer machine held eight B300 accelerators, with more memory than anything before them and arithmetic lean enough to push tokens in four-bit precision, and asked $53 an hour. The older machine held the same eight-GPU configuration built on the H200, the B300’s two-year-old predecessor, and asked $30.50. At first glance the premium looks not only justified but attractive to whoever owns the new machine. The B300 server costs about forty-three percent more than the old one, roughly $500,000 against $350,000, while its rent runs seventy-four percent higher. Per dollar of hardware the newer server produces more rent, so by the arithmetic most of the market runs, the B300 is the better business. The investor and operator Graeme Harrison chose a different denominator: the power the machines draw.¹ The new server pulls about 14.5 kilowatts at full load; the old one pulls about 7.7. Run each rent over each draw and the H200 earns $3.96 per kilowatt-hour of capacity while the B300 earns $3.66. The market is paying roughly eight percent more for the kilowatt that does less work. Harrison called the megawatt the true unit of compute economics. He is right, and following his inversion to the end shows where the durable money in artificial intelligence sits.

The Landlord’s Arithmetic

A data hall is a fixed envelope of energized power. Fill one megawatt with the old servers and you can host about 130 of them; fill it with the new ones and you can host 69. At the listed rents, the megawatt of old machines produces about $2.89 million a month and the megawatt of new ones about $2.67 million.² So every B300 a landlord installs occupies the power of almost two H200s, 14.5 kilowatts against 7.7, which sets a floor under the new machine’s rent: hosting it has to beat re-tenanting the same power with the machines it displaces, and the displaced rent runs about $57 an hour. Today the B300 asks $53. Economists call the principle the law of one price: the same input, sold in the same market, cannot carry two prices for long, and a kilowatt of energized capacity is the same input no matter which server draws it. The landlord is the arbitrageur. Every re-tenanting decision pushes the underpriced machine’s ask up and the overpriced machine’s rent down until the kilowatts earn alike. The tenant’s arithmetic sets the other boundary. A renter cares about cost per unit of work, so a tenant pays the B300 premium whenever the machine’s output advantage exceeds its price ratio, and for workloads built to exploit twice the memory and four-bit arithmetic it does: if the new server does twice the work, the B300 stays the cheaper way to produce output at any ask below $61. Landlord indifference puts a floor near $57; tenant indifference puts a ceiling higher still, wherever the productivity ratio clears the power ratio of 1.88. Today’s $53 ask sits below both boundaries, which is the full measure of the mispricing: the new machine rents below what the landlord requires and below what its own output justifies.

The correction has not happened yet because three frictions hold the window open. Most of the world’s models and pipelines are still tuned to the older architecture, so demand for the old machine today runs deeper than demand for the new one. The correction also needs landlords who can act, and most are locked on both legs of the trade: NVIDIA wound down Hopper-class production in favor of its next two generations, so no one can order new H200 megawatts to add more of the machine that pays better, and most existing halls cannot cool a 14.5-kilowatt server, so they cannot host the machine priced below parity either. A market whose participants cannot rebalance corrects at construction speed, in years, while listings reprice in weeks. And part of the old chip’s premium is the artifact of that same scarcity: a fixed and aging pool being cleared, which lends the rent a durability it does not have. One reseller’s sales page this month offers to quote “B300 or remaining Hopper inventory,” the artifact confessed in commercial language.³ Concede all of it. The trade between these two chips will close, through some blend of the new ask rising toward power parity and the old rent fading as the fleet ages, and the thesis expects exactly that. Windows like this open because buyers underwrite the newest-technology story while the arithmetic clears per kilowatt; they shut when the arithmetic wins. By the time you read this, the spread may already have narrowed.⁴ What survives the window is the unit the inversion revealed.

The Reserve, the Clock, and the Contract

Revenue per megawatt is the better lens, and still one line short of the full analysis. A megawatt of energized capacity is close to permanent. The silicon inside it is a wasting asset, and the waste is economic rather than physical: the chips still run in year four, but their rents decay as newer generations ship, which is why the companies that own the most of them keep arguing with their own auditors about useful lives.⁵ Keeping a megawatt earning at market rates therefore means replacing its tenant hardware roughly every three and a half years, and the honest way to carry that obligation is what real estate underwriting calls a replacement reserve, charged every year against the megawatt: fleet cost divided by economic life. Run it on the same figures. The megawatt of old machines holds about $45.5 million of hardware, so its reserve is about $13.0 million a year. The megawatt of new ones holds about $34.5 million, a reserve of about $9.9 million. The old configuration must set aside roughly $3.1 million more each year merely to remain what it is. The revenue advantage it buys, about $224,000 a month between the two megawatts, annualizes to roughly $2.7 million a year. Net of an honest reserve, the megawatt that wins on revenue loses on margin, and the inversion inverts again.

The replacement reserve is the zero-interest version of a tool corporate finance keeps for exactly this comparison, assets with different price tags and different lives: equivalent annual cost, the level yearly payment that would finance the machine over its life at your cost of capital, which is to say the hardware’s own implied rent. Run it at a twelve percent cost of capital, the middle of where rotating equipment borrows, and the numbers move further against the old megawatt: its hardware rents for about $16.7 million a year against $12.6 million for the new one, a gap of roughly $4 million versus $3.1 million on the zero-interest version, because the dearer fleet parks more capital while it waits to obsolesce.⁶ Two more filters finish the underwriting alongside the reserve. The first is the clock: the two lives are not the same three and a half years. The new fleet starts fresh, with its premium era ahead of it; the old fleet is deep into an architecture whose rents will erode within any holding period, so freezing today’s prices flatters it twice, once by skipping the reserve and once by treating a decaying rent as permanent. The second is the contract: a marketplace ask earns nothing until someone takes it, and spot demand arrives in lumps. Committed capacity trades at discounts of sixteen to thirty-nine percent to posted rates,⁷ and a two-year contract at $27 can be worth more than a $30.50 hope, because contracted cash flow can be reserved against, borrowed against, and underwritten. The metric that survives all three filters, the reserve, the clock, and the contract, is contracted contribution margin per megawatt, with utilization and duration attached, and with the reserve charged first.

The Moat Is in the Balance Sheet

Once the megawatt is the unit, cost of capital stops being one input to the purchase decision and becomes the decision itself. The structure contains two assets on two clocks: energized capacity that behaves like forty-year infrastructure, and silicon that turns over in less than four. Finance them as one thing and every refresh reprices the entire capital base at hardware risk, which is precisely how a market produces the mispricing in the opening scene. Finance them separately and the economics change character. Long-lived contracted power has borrowed in the range of six to eight percent for decades, because lenders can see forty years of output; rotating equipment prices like equipment, ten to fourteen percent, on terms matched to its life. Split the stack that way and contribution margin per megawatt stops swinging with each hardware generation, because the revenue duration that supports the cheap capital is anchored to the power contract rather than to whatever sits in the racks this cycle. Owning the power is the ticket that makes the structure available at all, since the cheap tranche exists only for whoever controls the megawatt and can contract its output; the balance sheet then decides how much of the rent the owner keeps. Heather Hall, a partner of mine, compressed the whole structure into a sentence I have not improved on: the chip is the trade; the power is the franchise. The distinction also dissolves the standard objection that tokens per dollar is what really matters. Per-token economics decide which tenant wins the hall. Per-megawatt economics decide what the hall earns. Confusing the two is how the newer tenant came to rent below the older one, kilowatt for kilowatt.

Ricardo in the Server Hall

In 1817 David Ricardo wrote down where rent comes from: it accrues to the factor that is scarce, productive, and impossible to reproduce at will.⁸ In his century that factor was fertile land, and the tenants who farmed it captured wages while the land captured rent. The interconnection queue for new grid capacity in the United States now runs in years while a GPU order runs in quarters, which tells you which factor cannot be reproduced at will. Energized capacity is the fertile land of this economy. The chips above it are tenants on a replacement chain, arriving every few years with better arithmetic and shorter leases, bidding against each other for the same scarce kilowatts. The inversion that opened this essay was the market briefly forgetting which party in that arrangement holds the durable claim, and the correction, when it completes, will be the market remembering. Claims on this land come in ranks: renting capacity in someone else’s hall is the weakest, holding your own grid interconnection is stronger, and generating your own power behind the meter is the strongest of all, because the operator who makes the megawatt stands in no queue.

Draw one line through this economy at the point where the megawatt is delivered, and the industry sorts itself. Below the line sit the things that endure: the generation, the interconnection, the energized hall, the forty-year assets, the capital that prices in single digits, the rent. Above it sits everything that merely plugs in: the servers, the architectures, the pipelines tuned to them, and, if the logic is carried honestly, the model companies renting the compute. The occupants rotate on three-and-a-half-year leases, each generation arriving with better arithmetic to bid for the same kilowatts, and the land collects from whoever moves in. The durable income statement of artificial intelligence is written per megawatt. Everything above the power line changes tenants.