CoreWeave, Inc. Class A Common Stock CRWV

Key data

The business

CoreWeave rents artificial-intelligence computing power. It buys Nvidia's newest GPUs by the rack, wires them into dense clusters with fast storage and networking, installs them in leased data centers, and sells time on them to companies training and running AI models. It reports as one operating segment, so there is no product-line breakout: the whole company is GPU capacity for hire, billed by the GPU-hour under multi-year contracts. The customers are a short list of the largest AI labs and cloud companies, which is both the strength and the danger, because a handful of contracts underwrite the entire build. The edge, such as it is, comes from being first and biggest: getting scarce chips before rivals, standing up clusters faster, and locking in power and contracts ahead of a crowd now chasing the same thing. It is a capital machine. Borrow, buy chips, fill them, repay, repeat, and the whole question is whether each turn of that wheel clears its cost.

The numbers

Growth is not the question here; the return on the borrowing is. Revenue has gone from $229 million in 2023 to a roughly $10 billion annual pace today, and last quarter alone did $2.58 billion.

The top line is compounding at a rate almost no company reaches, up 112% year over year and 24% sequentially, yet the losses got bigger, not smaller: $626 million lost in Q2 against $290 million a year earlier. The one bright spot is the operating line, where the loss narrowed to $49 million from $144 million in Q1, so the core business is closer to breakeven than the bottom line suggests. What sits between operating income and net loss is interest, the cost of the debt that bought the chips. On an adjusted basis the quarter beat, a $1.03 loss against a $1.21 estimate, the second beat in four prints after two misses.

The build is financed almost entirely with debt.

Long-term debt tripled in fifteen months, and of it $16.6 billion is senior notes carrying coupons as high as 9%. In 2025 the company spent $10.3 billion on property and equipment against $3.1 billion of operating cash flow, so it burned north of $7 billion in one year to build capacity; cash on hand was $5.5 billion at quarter end. The variant is narrow: the market is paying for the contracts to convert into cash faster than the chips depreciate and the coupons compound. The print that settles it is durable positive free cash flow, which the company has never produced and is not close to.

Management

The record is one-directional. Over the last twelve months insiders sold about $97.8 million across 56 transactions and bought nothing, not one open-market purchase. The selling runs to the top: CEO Michael Intrator sold $11.8 million on August 25 and another $10.7 million on September 1, with other officer sales alongside. Plan status is not disclosed in the filings I can see, so I cannot tell you how much is pre-scheduled 10b5-1 selling versus discretionary; treat it as unknown, not benign. Founders control the company through supervoting shares regardless. There is no profit for pay to reward yet, so compensation is set against growth and the stock price, not earnings.

How it fails or surprises you

The cost of money outruns the contracts. Debt reached $24.9 billion against roughly $10 billion of annual revenue, with senior notes at up to 9%. If GPU rental prices fall faster than the multi-year depreciation assumes, each cluster earns back less than it cost to finance. Watch quarterly interest expense against operating income; that gap was the whole loss this quarter.

A single customer walks. A short list of hyperscalers and labs underwrites the backlog. One renegotiation or cancellation strands capacity built on borrowed money, and the 10-Q's own risk language, plus New York's new 50-megawatt data-center moratorium, shows how fast the ground shifts. Watch for any disclosed change in a top customer's commitment or a step-down in contracted revenue.

Demand stays scarce and the wheel clears (right tail). If AI compute stays supply-constrained into 2027, utilization and pricing hold, and the operating turn already visible this quarter, from a $144 million loss to $49 million, compounds, free cash flow could inflect years before consensus pencils real earnings in 2029. The market prices none of that today. The first durable positive free-cash-flow quarter would signal the wheel is clearing.

Closing thoughts

The payoff unfolds over the next three years, as CoreWeave either converts contracts to cash faster than debt compounds, or doesn't. The distribution is genuinely two-tailed with fat ends. The left tail is a credit event: rental prices soften, a marquee customer trims, and a company with $24.9 billion of debt and $5.5 billion of cash cannot refinance on the same terms, which is how leveraged infrastructure businesses die. The right tail is that AI compute demand outruns supply for another two years, the contracts convert, and the operating turn already visible this quarter compounds into real cash. The concern flagged a day ago, that contracts may have been signed at prices that do not cover the cost of the money used to build them, neither broke nor cleared this quarter: the operating loss shrank, but debt kept climbing and free cash flow stayed deeply negative.

The bet is still that the biggest AI companies keep renting Nvidia chips from CoreWeave rather than building their own, and that CoreWeave earns back more than it borrows to serve them. What breaks it is the spread between what the contracts pay and what the debt costs, and the two numbers that tell you first are quarterly interest expense and free cash flow. Until free cash flow turns durably positive, this is a leveraged wager on demand staying scarce, priced for a 2029 that has to arrive on schedule.