CoreWeave Promises Profit in 2028. The Debt Says Otherwise

CoreWeave’s executives spent Tuesday telling investors the company will be profitable in the first half of 2028. The reassurance was necessary because the stock had fallen more than 40 per cent in a month, and because the argument about an AI bubble is no longer being conducted in the abstract.

Rows of GPU server racks in a data centre aisle under blue lighting
CoreWeave borrows to buy GPUs, builds the hall, then leases the capacity on long contracts to a handful of very large customers. Illustration: The Robot Belt

The top line is not the problem

Guidance for full-year 2026 revenue rose from USD 12.4 billion to a range of USD 13.0 billion to 13.3 billion. Third-quarter revenue reached USD 3.66 billion, up almost 170 per cent year on year and above the USD 3.55 billion consensus. The loss per share was 95 cents against an expected USD 1.16. On growth alone, the quarter was strong.

The company has never made a profit. Total debt stood at about USD 46.9 billion at the end of September, against USD 35.6 billion at the end of June, which means more than USD 10 billion was added in a single quarter. Capital expenditure jumped from USD 9.4 billion in the second quarter to USD 14.5 billion in the third. Gross margin in the second quarter was about 11.77 per cent, and net interest expense alone consumed USD 640 million. On top of the balance sheet debt sat roughly USD 16.3 billion of operating lease liabilities at the end of June, leverage parked off the reported debt line.

Backlog is not cash

Contract backlog reached USD 132.5 billion at the end of September, up from USD 104 billion disclosed at the end of June, after more than USD 25 billion of new customer commitments early in the third quarter. Roughly 41 per cent of the June backlog converts to reported revenue within twenty-four months.

The gap between backlog and cash is the whole business model. CoreWeave buys Nvidia GPUs on borrowed money, builds the halls, and leases capacity to Microsoft, OpenAI, Meta and Anthropic on long contracts. Chips, power, land and construction are paid up front. Revenue arrives on delivery and service milestones. The customers are also building their own capacity, which leaves open how much they will keep renting once it is finished.

The financing is the fault line

Some of CoreWeave’s 2031 junk bonds traded at yields near 16 per cent. In mid-October the yield on part of its unsecured debt climbed to about 15 per cent. On Tuesday the company said it would not borrow in the unsecured market at current yields, which is an admission that money is too expensive. About USD 9 billion of customer prepayments and deposits is due to arrive over the next three months, which buys time rather than solvency.

The history explains the intensity. CoreWeave started as an Ethereum miner in 2017, moved to cloud in 2020 and listed on Nasdaq in 2025. Nvidia is supplier, shareholder and backstop, having committed to take unsold capacity until 2032. Microsoft alone produced about 67 per cent of 2025 revenue, so a slowdown in one customer’s leasing ripples through the entire cash chain. Jensen Huang has publicly defended the company’s financial strength, and that endorsement is part of why it can keep borrowing while the market panics.

The short sellers are the cast of The Big Short. Seawolf’s founders compare CoreWeave to WeWork on the grounds of heavy upfront capital and permanent dependence on cheap funding, with one addition: GPU generations depreciate faster than drilling rigs. In mid-October the Philadelphia Semiconductor Index fell more than 11 per cent in a week, and peers including Nebius and Applied Digital fell with it. Goldman Sachs upgraded the stock to buy on 19 October, arguing the company can still raise. Gartner expects more than 80 per cent of these GPU-rental neoclouds to stop operating independently by 2030.

The stock closed up 5.83 per cent after hours at USD 51.57, still less than half its level near USD 100 at the end of the second quarter. To reach profit in 2028, CoreWeave has to cross annual spending of close to USD 40 billion and debt that may pass USD 50 billion, in a market that has already decided to charge it 15 per cent for the privilege.

The neocloud thesis was never about compute being scarce. It was about who could borrow the most against it, and the answer is now visible on the bond screen.

Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.ofweek.com/ai/2026-11/ART-201714-8120-30711322.html.

Leave a comment