CoreWeave has a $104 billion backlog — and a $640 million quarterly interest bill

CoreWeave’s AI data center business generated $2.58 billion in revenue last quarter — a 112% jump from a year earlier — and locked in $104 billion in signed future contracts from the AI labs that cannot build their own infrastructure fast enough. The numbers say one thing about the state of AI demand. The company’s balance sheet says something more complicated.

The customers are not ambiguous bets. Meta, which runs one of the world’s most heavily used AI systems, committed $21 billion to CoreWeave through 2032 — layered on top of a prior $14 billion commitment that makes the combined figure $35 billion from a single client. Anthropic, which builds Claude, signed a separate multi-year compute agreement. Quantitative trading firm Jane Street added $6 billion. The $104 billion total backlog — which does not include $25 billion in new commitments signed after the quarter closed — is not speculative pipeline. It is contracted revenue. Every Claude query, every Meta AI response, every model training run now flows through hardware that CoreWeave built, and is paying for, with borrowed money.

The underlying economics are unusual. On an adjusted operating basis, CoreWeave runs at 59% EBITDA margins — a level that most software businesses would consider extraordinary and that most hardware infrastructure companies cannot approach. But to build the physical clusters generating those margins, the company borrowed on a scale that now costs $640 million in interest every three months. That figure is guided to reach $860 to $940 million in the third quarter. For the full year, CoreWeave will spend between $35 and $39 billion in capital expenditure, the majority of it funded by debt.

CEO Mike Intrator pointed to momentum in managed inference — a newer service that lets AI developers run models on CoreWeave’s hardware without managing the infrastructure themselves. That segment grew from $1 million to $100 million in annual recurring revenue since launch, with a target of $250 million by year-end. New contracts, he noted, are being priced at contribution margins five to ten percentage points above recent quarters, driven by demand that is outpacing CoreWeave’s ability to build fast enough.

The tension in the results is arithmetical. Full-year adjusted operating income is guided at $960 million to $1.15 billion. Full-year interest expense, extrapolating from the Q2 result and Q3 guidance, will likely approach $3.5 billion. The company earns real money on its operations; it loses money once its lenders are paid. Whether that changes depends on whether AI demand compounds faster than the debt load does. Critics and short sellers have focused on this gap since CoreWeave’s IPO: at current trajectory, the company needs several more years of 112% revenue growth before interest stops dominating the loss.

For the workers and communities where CoreWeave’s 51 data centers operate — drawing 1.5 gigawatts of power across the United States, with contracted capacity at 4.2 gigawatts and a stated path to 8 gigawatts by 2030 — the business model’s durability is not an abstraction. Data centers are fixed capital. If AI demand slows, those buildings do not leave.

CoreWeave reports third-quarter results in November. By then, interest costs alone are expected to reach between $860 and $940 million for the quarter — more than the company’s projected adjusted operating income for the entire first half of the year.

Tags: Anthropic, meta, AI infrastructure, CoreWeave, CRWV, Jane Street