CoreWeave reports $2.58 billion quarter and $104 billion backlog as neoclouds build cash reserves
Understanding how GPU capacity is contracted, financed and priced explains why AI infrastructure looks nothing like the old public cloud.
The Next Platform's Timothy Prickett Morgan examined on 14 August the finances of the so-called neoclouds, the GPU-focused infrastructure providers that have grown alongside generative AI. The piece argues that these companies are accumulating cash and contracted backlog considerably faster than they are recognising revenue.
CoreWeave, the largest of the group, reported second-quarter 2026 revenue of $2.58 billion, a 2.1-fold increase. It posted an operating loss of $49 million and a net loss of $626 million, spent $9.4 billion on capital expenditure in the quarter, and ended with $6.41 billion in cash and equivalents. Its revenue backlog stood at $104.2 billion at the end of June, and the company has since added another $25 billion. CoreWeave runs 51 data centres with 1,500 megawatts of active power, which the article estimates at roughly 1.1 million GPUs, and has 3,700 megawatts under contract.
Its customer list now includes companies that once looked like competitors. Long-term agreements cited include $22.4 billion with OpenAI and $32 billion with Meta Platforms, along with deals with Google and Anthropic, plus a $1 billion strategic investment from Jane Street.
Nebius, the second company profiled, reported close to $6 billion in deferred revenue and more than $40 billion in customer capacity commitments, with a target of 5 gigawatts of contracted power by the end of 2026. Nebius told investors it sees a pricing opportunity of $40 million to $50 million per megawatt for short-term contracts of three to six months, although long-term deals remain its core business. The article notes that many neoclouds began in cryptocurrency mining and initially undercut established cloud providers on GPU pricing, and that with demand far above supply, terms come down to who holds GPU allocations.
Why it matters
Compute demand still exceeds supply, so whoever holds GPU allocations and power sets the terms. Model builders signing multi-year, multi-billion-dollar commitments with specialist providers means the largest AI workloads no longer land automatically with the three big hyperscalers. Backlog and prepayments now function as a financing instrument for data-centre construction, which ties AI progress to credit markets in a new way. The capacity unit that matters has shifted from the virtual machine to the megawatt.
Compute demand still exceeds supply, so whoever holds GPU allocations and power sets the terms.
- Quarterly revenue$2.58B
- Cash and equivalents$6.41B
- Quarterly capital expenditure$9.4B
- Revenue backlog at end of June$104.2B
Figures: CoreWeave figures as reported by The Next Platform
What you can learn from this
- Backlog is a promise; revenue is a delivery. Under accrual accounting a company recognises revenue as it provides the service, so a five-year contract worth billions appears in reported revenue only in small slices each quarter. Backlog, sometimes called remaining performance obligations, is the total contracted value still to be delivered. Deferred revenue is cash already collected for service not yet provided and sits on the balance sheet as a liability. A large backlog beside modest revenue is exactly what a business that sells capacity years ahead of building it looks like.
- GPU clouds sell reserved clusters, not spot instances. Traditional public cloud earns much of its money from on-demand instances billed by the second, which suits variable workloads. Training a large model needs thousands of accelerators for months, connected by high-bandwidth networking that cannot be assembled on demand. Providers therefore sell whole clusters on multi-year reservations, which locks in utilisation for the seller and locks in scarce supply for the buyer. Short-term contracts command a premium because they shift risk back to the provider.
- Power is the binding constraint, so capacity is quoted in megawatts. A modern accelerator rack can draw over 100 kilowatts, and a site's grid connection sets a hard ceiling on how many racks can run. Contracted power is capacity a utility has committed to deliver; active power is what is energised today. Because grid connections take years to secure, the gap between the two figures is a rough map of how much future build-out is already locked in.
- Capital intensity changes the risk profile. Spending $9.4 billion in a quarter to earn $2.58 billion is only viable when the hardware is expected to earn revenue over several years. The provider carries the risk that GPU generations turn over faster than contracts amortise, or that customers renegotiate. A net loss far larger than the operating loss usually points to interest and financing charges from the debt that funded the hardware.
- Customers can also be competitors. When large model builders rent from neoclouds, the boundary between cloud provider and cloud customer blurs. The neocloud gains anchor tenants that make its debt easier to raise, while the tenant gains capacity it could not build in time. Concentration cuts both ways: a handful of contracts make up most of the backlog, so one renegotiation can move the whole picture.
We teach this
How to use this in practice
- Price a training job three ways. Take a concrete workload, for example fine-tuning a 7-billion-parameter model on 64 GPUs for one week. Look up on-demand, one-year reserved and spot pricing on any two providers' public price lists and compute the total for each. Include networking and storage line items if the price list shows them separately, since cluster interconnect is often priced apart from the accelerators. Done looks like a small table showing the cost spread, with a note on which option you would choose and why.
- Read one quarterly report's backlog note. Download the latest earnings release or quarterly filing from a listed infrastructure provider and find the figures for revenue, remaining performance obligations, deferred revenue and capital expenditure. Write each in a four-cell grid and calculate the ratio of backlog to quarterly revenue. Done means you can explain in two sentences why the four numbers can diverge so widely.
- Draw the megawatt-to-GPU chain. Sketch the steps: utility feed, substation, facility power, rack power, and GPUs per rack, with a rough number on each, then work backwards from a figure such as 1,500 megawatts to estimate accelerator count. Compare your estimate with the article's 1.1 million figure. Then add a second column with numbers for any small setup you know, even a single rack, to see how the ratios change at small scale. Done is a diagram with every assumption labelled so someone can argue with it.
- Check how your own cloud spend is committed. Open the billing console for whatever cloud account you use and identify what share of last month's spend was on-demand versus reserved or savings-plan pricing, and whether any GPU instances appear. If you find none, write that down too; it means your exposure to GPU pricing is indirect, through the model APIs you call. Done looks like a single percentage for committed spend and a list of GPU line items with their pricing model beside each one.
Sources
- The Neoclouds Build Cash Hoards Faster Than Revenues — The Next Platform
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report