CoreWeave's $104B Backlog Is a Promise to Deliver, Not Cash in the Bank

An AI cloud company reported approximately $104.2 billion in future revenue backlog as of June 30, 2026.

For CoreWeave, that enormous number reflects customers willing to commit to computing capacity for years ahead. But a backlog is not a deposit account, and it is not a pile of GPUs already earning their keep.

The headline conceals a physical sequence: a customer signs for future capacity, an infrastructure company secures electricity and buildings, equipment is installed and tested, and only then can a service be delivered and revenue recognized.

The gap between those steps is where CoreWeave's opportunity—and its financial risk—lives.

What the $104.2 billion figure includes

CoreWeave defines its revenue backlog to include remaining performance obligations plus certain other estimated future customer-contract revenue. The company says these amounts are subject to service delivery and availability conditions.

This is different from cash already collected, a forecast of net profit or recognized revenue from the current quarter.

A customer might contract to use a cluster over several years. CoreWeave earns the relevant revenue as it performs under the agreement, subject to the contractual terms. A contract covering future capacity can be commercially valuable even when the company has not yet installed all the equipment needed to supply it.

The company also reported more than $25 billion in additional net new commitments early in the following quarter. Those commitments were explicitly outside the June 30 backlog figure.

Adding the two numbers without their dates and definitions would turn a complex disclosure into a misleading headline.

Why backlog needs a power station

A signed computing contract cannot cool a processor or connect a building to a substation.

Large AI deployments require electricity, networking, high-density cooling, accelerators, memory and specialized engineering. The facility must be brought online before a customer can reliably consume its contracted capacity.

CoreWeave reported approximately 1.5 gigawatts of active power around the end of the second quarter of 2026, while its contracted power footprint was larger and included future development.

Active and contracted power are not additive categories. One is an operating-state measure; the other describes a pipeline of present and planned resources.

The difference matters because data-center construction runs on a separate calendar from commercial contract signing. Delays in grid connections, cooling hardware, construction or GPU deliveries can postpone when a valuable obligation becomes billable capacity.

This is the part of the AI boom that looks less like software and more like an industrial development program.

Revenue growth has a financing cost

In the second quarter of 2026, CoreWeave reported approximately $2.575 billion in revenue, up sharply from the prior-year quarter. It also recorded a GAAP net loss of about $626 million and approximately $640 million in net interest expense.

Interest is already reflected in net earnings; it should not be added to the reported net loss as though it were a separate incremental loss.

The company also described substantial capital investment in its data-center expansion, including roughly $9.4 billion in quarterly capital expenditures, and management projected major additional spending over the rest of the year.

These figures illustrate the timing challenge. Computing capacity is exceptionally expensive to create. Much of the spending happens before all of the contracted service is recognized as revenue.

A fast-growing order book can make this spending rational, but it does not magically finance itself.

Why adjusted EBITDA cannot settle the debate

CoreWeave reported approximately $1.51 billion in adjusted EBITDA for the quarter. That can be useful for analyzing aspects of operating performance, but it is a non-GAAP measure.

Adjusted EBITDA is not the same as net earnings or free cash flow. It does not independently show how much money remains after capital expenditures and all financing obligations.

A complete picture requires keeping four separate questions straight:

Demand: How much future computing have customers contracted for?

Delivery: How much installed, tested capacity can be made available on schedule?

Operations: How much revenue is recognized and what does it cost to run the service?

Financing: How much debt, interest and up-front capital does the build-out require?

Confusing any of these measures can produce a story that is either unrealistically euphoric or unfairly catastrophic.

The landlord has a different balance sheet

Some of the physical infrastructure supporting CoreWeave belongs to other companies.

Core Scientific, for example, disclosed major colocation commitments to CoreWeave. That company finances and operates its own infrastructure arrangements, and it recognizes revenue under its own agreements as power-ready capacity is delivered.

This creates a chain of dependencies: the AI compute customer needs CoreWeave; CoreWeave needs infrastructure; the landlord needs construction progress and a paying customer. The companies may benefit from longer-term contractual protections, but their assets, liabilities and risks should not be collapsed into one set of financial statements.

A landlord's secured borrowing cannot simply be described as CoreWeave's corporate debt. Nor does CoreWeave's customer backlog mean every upstream landlord has already earned its future fees.

The same demand can support several layers of business without every layer being equally liquid.

The downside is not necessarily that demand disappears

There are genuine protections in committed cloud arrangements: capacity reservations, minimum payments and negotiated terms can improve revenue visibility and help operators obtain financing.

But customers still care about hardware generations, cost per unit of computation, performance and delivery dates. A cloud provider must keep its expensive systems attractive while newer processors and more efficient software arrive.

An operator can have real customer demand and still be challenged by timing, debt costs or the speed at which usable capacity ramps.

This does not mean CoreWeave's business is fictitious or that a default is inevitable. The company's disclosures show large commercial opportunity paired with large capital requirements.

Watch the conversion, not only the contract

Over time, the most informative evidence will be how much backlog becomes delivered computing, recognized revenue and sustainable cash generation.

That requires active facilities, satisfied customers, controlled investment costs and financing that can withstand delays.

The $104.2 billion figure is a statement about demand under agreements.

The real industrial achievement will be converting those agreements into working computers—and keeping enough of the economics when the bills arrive.