Self-hosting's total cost worked out what a machine costs to run: power, cooling, colocation, and the engineer nobody budgets for. This chapter is about the decision that comes before that one, and it turns on a single change of shape.
Rent a GPU and cost is variable. Stop the instance and the meter stops. Buy the same GPU and cost becomes fixed: the depreciation charge lands every month at the same size whether the card is saturated or dark. Nothing else about the hardware changes. The invoice does.
That conversion is the whole decision, and it has a number attached — the fraction of the time you must actually be using the thing before ownership wins. Call it the utilisation floor. Almost every self-hosting model I have seen omits it, because the alternative it compares against is a rental bill computed at 100% utilisation, which is a bill nobody who rents ever pays.
Someone else already ran the experiment, and had to publish it
The problem with sizing this from the outside is that the inputs are private. NVIDIA publishes no DGX list price. Colocation providers quote rather than publish. So the honest move is to find someone who bought at scale and is legally obliged to disclose what happened.
CoreWeave is that company. It is a pure-play GPU cloud — it does not sell search ads or retail alongside, so its capital spending is not diluted by other businesses — and since its 2025 listing it files with the SEC. Its FY2025 Form 10-K discloses both halves of the ratio nobody else publishes: what it spent, and how much capacity that bought.
| Year end | Data centres | Active power | Capex | Revenue |
|---|---|---|---|---|
| 2023 | 10 | ~70 MW | $2,943M | $229M |
| 2024 | 32 | ~360 MW | $8,702M | $1,915M |
| 2025 | 43 | over 850 MW | $10,309M | $5,131M |
Three years of capital spending sum to $21,954 million against over 850 MW of live capacity at the end of 2025. That is $25.83 million of capex per megawatt — and because the megawatt figure is disclosed as a floor rather than a point, every per-megawatt number in this chapter is an upper bound on the true one.
The marginal years are more informative than the cumulative figure, because they show the direction:
| Capacity added | Capex that year | $ per MW added |
|---|---|---|
| 2024: 290 MW | $8,702M | $30.01 million |
| 2025: 490 MW | $10,309M | $21.04 million |
Cost per megawatt fell about a third year over year. Some of that is genuine — better siting, denser racks, buying power. Some of it is an artefact of the lag between spending money and lighting capacity, which I come back to below. Either way the order of magnitude holds: standing up a megawatt of GPU capacity costs tens of millions of dollars, and the number is now sourced rather than guessed.
What a megawatt returns
The same filing gives the other side. 2025 revenue of $5,131 million against over 850 MW is $6.04 million per megawatt-year.
That understates productivity, because capacity grew through the year and the revenue was earned against a smaller average fleet. Against the average of the opening and closing figures — 605 MW — it is $8.48 million per megawatt-year. I use the year-end basis for the rest of this chapter because it is the conservative one and because it matches the basis of the capex figure, but the average-basis number is the fairer read of how hard the machines actually worked.
Now put the two together at the life the company itself assigns to the equipment.
The six-year life, and who moved which way
CoreWeave depreciates technology equipment over six years. That is not an inference — the useful-life table in the 10-K says so.
More interesting is that it used to say five. The filing discloses that the company "changed its estimate of the useful life for its computing equipment utilized in data centers from five to six years, reflecting continuous advancements in hardware performance, software optimization, and data center design improvements." The change reduced FY2023 expenses by $20 million and added $0.10 per share.
Set that beside the disclosure in self-hosting's total cost: Amazon moved its server life the other way, from six years to five, citing artificial intelligence, and took a $1.4 billion increase in depreciation and a $1.0 billion reduction in net income for it.
The two companies most exposed to GPU economics moved their depreciation schedules in opposite directions, within a couple of years of each other, and both cited technology as the reason. One decided its machines last longer than it thought; the other decided they last less. Both statements are audited. Neither is wrong. They are answers to different questions — Amazon is writing down a fleet it must keep competitive across every workload, CoreWeave is writing down GPUs it re-lets on multi-year contracts — but the disagreement is the point. If two audited filings cannot agree whether a server lasts five years or six, the depreciation schedule in your own model is a judgment call, and you should label it as one.
That is a 20% swing in the largest fixed cost of ownership, decided by opinion.
The utilisation floor
Take the six-year life at face value. $25.83 million of capex per megawatt, straight-lined over six years, is $4.30 million per megawatt-year of depreciation. Against $6.04 million of revenue per megawatt-year:
Depreciation alone consumes 71.3% of revenue.
Before power. Before cooling. Before staff, interest, land, or the sales team. On the one cost that is completely insensitive to whether the machine is doing anything.
Read from the other direction, that ratio is the utilisation floor. If revenue scales with how busy the fleet is, then a fleet running at 71.3% of its current utilisation earns exactly its depreciation and nothing else. Everything above that line pays for the rest of the business. Everything below it is a loss that arrives on schedule whether or not a single token is generated.
This is why the question to ask about a purchase is never "is the hourly rate cheaper than renting" — at full utilisation it almost always is. The question is:
What fraction of the next six years will this machine be busy, and am I confident enough in that number to convert it into a fixed cost?
For a research team whose GPUs idle overnight, at weekends, and between projects, real utilisation is often under a third. At that level the fixed charge per useful hour is triple the headline, and the rental comparison inverts.
What buying ahead looks like on a balance sheet
Two more disclosures show the commitment ownership actually is.
CoreWeave's 2025 capex of $10,309 million was 2.01× its revenue of $5,131 million that year. Cumulatively, $21,954 million of capex against $7,275 million of revenue is 3.02×. The company spent three dollars building for every dollar it billed.
And it is not finished. Against over 850 MW of active power at year end, total contracted power capacity was approximately 3.1 GW — 3.65× what is live, and disclosed as capacity the company "expect[s] to deploy over future periods."
What makes that survivable rather than reckless is the other side of the book: $60.7 billion of remaining performance obligations at year end, up from $15.1 billion — a 4.02× increase — at a weighted-average contract duration of about five years. That is 11.8 years of 2025 revenue already under contract.
That is the actual shape of the buy decision at scale: capex ahead of revenue, sold forward on contracts roughly as long as the depreciation schedule. Ownership works when demand is contracted for about as long as the asset is written down over. It is the matching of those two durations, not the hourly price, that makes the arithmetic close. If your own demand is not contracted — if it is a roadmap, or a hope — you are taking the fixed cost without the thing that justifies it.
The lag, stated plainly
One caveat matters enough to state rather than bury. Capex in a year does not all become active power in that year. Money spent in December 2025 lights capacity in 2026. So the marginal $/MW figures above mix a numerator and a denominator from slightly different periods, and the true cost of a delivered megawatt is somewhat lower than $30.01 million in 2024 and somewhat higher than $21.04 million in 2025.
The cumulative figure of $25.83 million per megawatt is the more robust of the three, and it is still an upper bound: it charges three years of spending — including work in progress toward the 3.1 GW contracted — against only the capacity that was live on 31 December 2025. Treat it as a ceiling on what a live megawatt costs, not an estimate of it.
The diagnostic
Four questions, in order. The first one decides the rest.
- What is your honest utilisation over the depreciation period, not your peak? Measure it on what you run today. If you cannot measure it, you are not ready to buy.
- Is that number above the floor? Fixed annual cost divided by annual revenue — or by the rental bill you would otherwise pay at that same utilisation. Below the line, renting wins regardless of the sticker price.
- How long is your demand contracted for, and how does that compare with your depreciation schedule? The closer those two durations, the safer the purchase. A six-year write-down against six-month commitments is a bet, not a plan.
- Which life did you assume, and would the decision survive the other one? Five years versus six is a 20% swing in the dominant cost. Two audited filers disagree. If your answer flips between them, you do not have a decision — you have a preference.
What this chapter is not saying
Not that owning is wrong. CoreWeave's disclosures describe a business that raised capital specifically to do this, contracted its output years forward, and is compensated for the risk. That is a coherent strategy at that scale.
It is saying that the numbers people use to justify smaller purchases are usually the rental comparison at full utilisation and the ownership comparison at zero risk, and that both halves of that are wrong in the same direction. The disclosed figures give you a floor to check yourself against. Use them as a ceiling on the upside and a check on the utilisation you are quietly assuming.