Summary

  • A customer can reserve computing for a few months and leave when the job is done. CoreWeave keeps the servers. Its higher prices for short bookings are welcome, but the next customer still matters.
  • The company spreads equipment costs over six years in its accounts. Earning that money back is a different job: deliver the promised service, find more paid work and cover the bills along the way.

Consider an AI team renting a cluster for a three-month training run. If the work finishes on time, the team can leave with its model rather than a roomful of machines to maintain. It has bought a service, not a data centre. The owner of the servers has a different finish line: those machines still need to earn back what they cost.

This is the attraction, and the hard work, of CoreWeave's business. Customers can buy computing without owning all the equipment. CoreWeave can sell the equipment's time again. A high price for the first booking helps, but a gap before the next booking earns nothing. For the customer, meanwhile, a reservation is only useful if the promised computing is ready when the work needs to start.

CoreWeave's September announcement puts a price on some of that business. On 17 September 2026, the company said it had signed customer contracts during the third quarter lasting around three to six months, at higher prices. Comparing them with longer bookings needs a common yardstick. A daily car-hire rate serves a similar purpose: it lets a buyer compare prices, but does not say how many days the car has been booked.

CoreWeave uses a yearly rate. It calculates what the short contracts would earn over a year at the same pace, then divides that amount by the power needed for the relevant server clusters. Its reported figure was approximately $40 million per megawatt. That lets it compare revenue against the scale of the computing operation. It does not turn a three-month booking into a year's sale. Company announcement.

Correction, 21 October 2026: An earlier version incorrectly described the announcement as lacking a price unit and contract duration, and left open whether CoreWeave was buying or selling capacity. It concerns customer sales. We have also corrected the treatment of disclosed equipment lives, contractual terms and Amazon's historical depreciation change, and replaced secondary references with the relevant primary documents. The original publication date is unchanged.

A year's rate is not a year's booking

Annualizing a short contract makes rates easier to compare. It does not oblige the customer to stay for a year. Nor is a dollar-per-MW figure a GPU-hour price that a buyer can place directly beside another supplier's quote: hardware, service configuration and the work completed still matter.

The announcement is evidence of CoreWeave's own recent sales, not an independent price survey or a repricing of its entire customer book. A strong short-term rate can help equipment earn its keep. To establish lasting pricing power, the next booking matters as well: whether it arrives, how much it pays and what it costs to serve.

Short contracts are not the whole company. The annual report describes committed service agreements with initial terms generally ranging from one to six years, usually requiring payment whether or not the customer uses all the reserved capacity. Revenue from those arrangements is recognized over the service period. On-demand services are recognized as consumed; a short fixed commitment should not automatically be called on-demand. Prepayments initially sit in deferred revenue, not in earned sales. 2025 annual report, revenue recognition.

This matters to both sides of a reservation. A customer may value assured access more than the option to leave. CoreWeave gains a firmer claim on future payments, but still has to supply the service. Neither a signature nor an advance payment switches on an unfinished cluster.

The large number has a timetable

There is already a large queue of work to deliver. CoreWeave's accounts put a value on services customers have committed to buy but the company has not yet provided. Accountants call this remaining performance obligations, or RPO. At 30 June 2026, the figure was $103.7 billion.

CoreWeave expected 41% to become revenue over the next two years, ending in June 2028. Another 39% was expected in months 25–48, and the rest in months 49–78. The dates matter because a large signed commitment is not yet a completed service. They describe when the company expects to count the work as revenue, not when every customer pays or every individual contract ends. Second-quarter filing, revenue note.

The distinction gives the number some useful shape. It also stops the short contracts in September from standing in for the existing business. RPO includes billed and unbilled commitments and allows for estimated adjustments such as service credits and delivery delays. It cannot settle, on its own, whether a particular site will be ready on the date a customer needs it.

The customer base is another part of that timetable. Three unnamed customers accounted for 36%, 26% and 10% of second-quarter revenue. Those are revenue shares, not shares of RPO. They nevertheless explain why delivery and renewal decisions by a few buyers can matter far more than the number of contracts suggests.

Six years is an estimate, not a warranty

CoreWeave's equipment assumption is not hidden. Its annual report assigns technology equipment a six-year estimated useful life and uses straight-line depreciation. The company moved computing equipment from five to six years in January 2023; it disclosed a $20 million reduction in total expenses for that year from the change. The second-quarter 2026 filing reports no material change to those accounting policies.

Stretching the allocation of a purchase cost changes when the expense appears. It does not change the invoice already paid, make a processor faster or secure another customer. Equally, a six-year estimate is not evidence that the accounts are wrong. Equipment can move to less demanding work and continue earning. Whether it does so profitably is an operating question.

There is a history of such revisions beyond CoreWeave. Amazon extended server lives from four to five years and network equipment from five to six effective January 2022. Microsoft moved both categories from four to six years from fiscal 2023. Those are historical company-specific decisions, not a shared warranty for today's GPU fleet. Amazon's 2022 filing; Microsoft's 2023 report.

The useful comparison is therefore not simply “six-year asset, six-year contract.” Matching lengths do not prove that revenue covers the purchase, electricity, operating costs and financing. A shorter contract need not be a problem if subsequent work pays well. A longer one need not be attractive if delivering it costs too much.

The next booking is where the estimate meets the business

For a customer, the practical exposure is delayed or inadequate service. For CoreWeave, it includes equipment that costs money to run but may command less at renewal. For a lender, repayment also depends on when payments fall due and which assets or guarantees support the debt. These risks overlap without being interchangeable.

What can a lender claim if repayment goes wrong? The answer depends on the loan. CoreWeave distinguishes borrowing with and without recourse beyond specified assets or borrowers. One arrangement, called DDTL 4.0, has particular security and limited guarantee exceptions. It does not describe every company loan. Customers have a different set of protections: the annual report describes refund exceptions for an uncured material breach, bankruptcy or insolvency. A right to seek money back is important, but it is not the same as having the money available when a service fails.

If equipment stops earning enough, its recorded value may need to be written down. That is an impairment. A disappointing renewal does not automatically require one: CoreWeave's accounting policy calls for a test of whether the recorded value can be recovered when circumstances warrant it, followed by a fair-value measurement if a write-down is needed. Retiring a machine, changing its estimated working life and writing down its value are different decisions.

The next useful evidence would connect the pieces: delivery progress, realized revenue, renewal rates and the work performed by older equipment. September's price measure makes the first booking easier to see. It does not tell us who will pay for the next one.