CoreWeave (NASDAQ: CRWV) delivered a second quarter that lays bare the central tension of the AI infrastructure boom. Demand is accelerating faster than the company can build. But building is expensive.
Revenue hit $2.575 billion, up 112% from $1.212 billion a year ago and modestly ahead of the $2.556 billion consensus estimate. Adjusted EBITDA more than doubled to $1.51 billion, a 59% margin that remains best-in-class for a cloud infrastructure business at this scale. The headline beat, however, masks a widening GAAP net loss of $626 million, or $1.14 per share, compared to a $290 million loss a year ago.
The divergence between GAAP and non-GAAP metrics is the quarter's defining feature. GAAP operating income swung from a $19 million profit in Q2 2025 to a $49 million loss, driven by $1.393 billion in depreciation and amortization and $640 million in net interest expense. Those aren't signs of operational weakness. They are the cost of building a purpose-built AI cloud from scratch at hyperscale velocity.
CoreWeave expanded active power capacity by nearly 500 megawatts to 1.5 GW during the quarter, with total contracted power reaching approximately 3.7 GW. Property and equipment, net, ballooned to $46.7 billion from $30.6 billion at year-end 2025. The cash flow statement tells the story plainly: $6.4 billion in capital expenditures during the quarter, funded by $10.1 billion in financing proceeds, including a $3.1 billion term loan and over $10 billion in unsecured debt and convertible bonds raised during the period.
The revenue backlog of approximately $104 billion as of June 30 is the single most important number in the release. It represents remaining performance obligations plus estimated future revenue under committed customer contracts, subject to delivery and availability conditions. CoreWeave disclosed an additional $25 billion in net new customer commitments added in early Q3, suggesting the pipeline is accelerating, not plateauing. For context, that backlog is roughly 10x trailing twelve-month revenue, a coverage ratio that implies multi-year visibility unusual even for infrastructure-as-a-service businesses.
Customer wins spanned AI labs, hyperscalers, and enterprises, including Bentley Systems, Caterpillar, Grammarly, and Isomorphic Labs, alongside expanded relationships with Databricks, Hudson River Trading, and Runway ML. The company also completed the industry's first bring-up and validation of NVIDIA's Vera Rubin NVL72 platform, positioning it to capture the next generation of GPU demand.
The adjusted EBITDA margin slipped to 59% from 62% a year ago, a 300bp compression that reflects the mix shift toward lower-margin colocation and power arrangements as the company scales its physical footprint. That's a structural feature of the model at this stage, not a cyclical one. As more revenue comes from contracted power and data center services rather than pure GPU compute, gross margins will compress before operating leverage kicks in at scale. The adjusted operating income margin tells a similar story, falling to 5% from 16% a year ago, as stock-based compensation and acquisition-related costs added $166 million in non-GAAP adjustments.
CoreWeave did not provide forward guidance in the release, stating it would do so on the earnings call. The absence of guidance, combined with the $25 billion in early Q3 commitments disclosed outside the backlog figure, suggests management is comfortable with the demand trajectory but may be managing expectations around the pace of margin recovery.
The balance sheet ended the quarter with $5.5 billion in cash and equivalents, against $31.4 billion in total recourse debt. Net leverage is high, but the cash flow profile is improving: operating cash flow turned positive at $679 million for the quarter, compared to negative $251 million a year ago, driven by $1.365 billion in deferred revenue additions.
The real question for CoreWeave is not whether demand exists. The $104 billion backlog and accelerating enterprise adoption answer that decisively. The question is whether the company can convert that backlog into GAAP profitability before the interest bill and depreciation schedule overwhelm the operating leverage. Adjusted EBITDA suggests the underlying business is healthy. GAAP net income will take longer to arrive, and the path depends on how quickly the installed base of GPUs and data centers starts generating cash rather than consuming it.