Last week, one series of CoreWeave’s unsecured notes traded at a yield above 15%. Market participants were already treating the level as distressed. The shares fell roughly 16%, to the mid-$60s, while credit default swaps traded above 800 basis points.
At a 15% yield, the market is no longer questioning CoreWeave’s ability to grow. It is questioning whether the company can turn its backlog into cash before the cost of financing that growth overwhelms the returns.
This is no longer an isolated move in one company’s debt. Bonds financing several of the largest US data-centre projects have also come under pressure. The market is starting to distinguish between the demand for AI computing capacity and the ability of the infrastructure built to supply it to generate adequate returns.
The wider repricing of AI infrastructure
The bond market has begun to challenge the assumption that the revenues generated by the AI boom will be sufficient to support the enormous amounts of debt being accumulated to finance it.
The debt financing the Hyperion data-centre project, backed by Blue Owl and Meta, fell from above par in early July to below 92 cents on the dollar. Bonds linked to two projects developed by Hut 8 traded at around 94 and 92 cents, respectively. Debt financing Blackstone-owned QTS’s Project Excalibur, leased to Microsoft, fell below 88 cents.
Oracle-related projects have also experienced pressure. Approximately $18bn of loans associated with Project Jupiter in New Mexico reportedly entered stressed territory following delays involving the site’s power supply and local opposition. Bonds financing Oracle’s Michigan AI campus fell to around 87 cents on the dollar.
These projects are not interchangeable credit exposures. Their tenants, financing arrangements, collateral packages and remaining construction risks differ. But the direction of travel is difficult to dismiss.
The repricing has several potential explanations. Investors are questioning the timing of AI monetization, the scale of the capital expenditure required and the execution risks attached to power, construction and permitting. OpenAI’s decision to defer its IPO until next year and reports of slower quarterly sales growth at Anthropic have added to those concerns.
How much capital can the industry commit before the revenues generated by that infrastructure catch up?
Gartner has estimated that data-centre owners would need to generate approximately $8.5 trillion in revenue by the end of 2030 to cover their capital and financing costs while achieving healthy returns. Its forecast for consumer and corporate spending on AI services that would flow through to data-center providers was no more than $3.6 trillion. Gartner has warned that around 80% of neo-cloud providers could cease independent operations by 2030 and that more than $4 trillion in asset value could be exposed to write-downs through 2032 if infrastructure investment continues on its current trajectory.
These are forecasts, not certainties. But the scale of the potential gap explains why the market is reassessing the assumption that almost any available AI capacity will eventually generate an adequate return.
The risks extend beyond demand. Political opposition to data centres is growing in the US, with concerns about electricity consumption, water, noise and local pollution delaying or complicating projects. Permitting and power constraints can postpone revenue while financing costs continue to accrue. The Bank for International Settlements has also highlighted the interconnected nature of some AI financing and commercial relationships, in which companies invest in or finance counterparties that are also their customers or suppliers.
Taken together, these developments point to a more discriminating credit market, questioning which projects can be completed on time, which operators can fund them without continually returning to the capital markets, and which assets will retain their value if demand or pricing disappoints.
CoreWeave brings these questions into particularly sharp focus.
CoreWeave began in September 2017 as Atlantic Crypto, a small Ethereum-mining operation in New Jersey. It pivoted towards renting Nvidia graphics processors by the hour in 2019 and 2020 and eventually adopted the CoreWeave name. The business had little non-crypto revenue before 2022, but it was well positioned to benefit from the subsequent shortage of accelerated computing capacity.
The expansion has been rapid. By August 2026, CoreWeave was operating around 1.5 gigawatts across 51 data centers, with contracted power capacity of approximately 4.2 gigawatts.
The company listed in March 2025 at $40 a share, raising approximately $1.4 billion. Nvidia, already an investor, subsequently made a further $2 billion equity investment in January 2026. Nvidia has a commercial interest in the expansion of AI computing capacity, and its involvement may help CoreWeave secure equipment and maintain access to capital.
But a strong strategic backer does not remove the underlying financing risk.
The model is straightforward to describe and demanding to finance. CoreWeave signs multiyear, take-or-pay contracts for computing capacity, acquires processors and builds the infrastructure needed to deliver that capacity. It then uses a combination of corporate debt, asset-backed facilities and project financing to fund further expansion.
The contracted revenue provides visibility. The challenge is that much of the cash needed to deliver it must be spent before the revenue arrives.
The scale of the commitments is substantial. CoreWeave reported remaining performance obligations of $60.7 billion in December 2025. Its reported revenue backlog reached $105 billion in June 2026. The company added more than $25 billion of net new commitments early in the third quarter, which were not included in that June figure.
But backlog is not cash. It must be converted into installed capacity, recognized revenue and ultimately cash available for debt service. Each step depends on execution, equipment availability, electricity supply, customer requirements and the economics of the contracts themselves.
Customer concentration adds another layer of risk. The largest customer’s share of revenue reportedly declined from around 85% towards 35% during 2025 as Meta and OpenAI expanded their relationships with CoreWeave. That diversification is positive, but the business remains dependent on a relatively small group of very large customers.
Impressive growth, but the cash-flow burden is substantial
CoreWeave’s operating performance has been exceptional by conventional infrastructure standards.
Revenue increased from $229 million in 2023 to $1.9 billion in 2024 and $5.13 billion in 2025. In the second quarter of 2026, revenue reached $2.575 billion, up 112% year on year. Adjusted EBITDA was $1.51billion, equivalent to a margin of approximately 59%.
Those are powerful growth and margin figures. However, they need to be considered alongside interest expense, capital expenditure and the cash required to keep building capacity.
The second-quarter operating loss was $49 million. Net interest expense was $640 million, contributing to a net loss of $626 million. Interest expense for the first half reached $1.176 billion.
Capital expenditure is the other major pressure point. CoreWeave spent approximately $9.4 billion in the second quarter and has guided to full-year capital expenditure of $35 billion to $39 billion.
The result is a business whose adjusted EBITDA margin looks attractive while its funding requirement remains enormous, generating free cash burn at roughly $10 billion for the first half, depending on the treatment of working capital and other cash-flow items.
CoreWeave is growing quickly, and its contracted capacity offers a route to much greater future earnings. But operating cash generation currently covers only a fraction of the investment program.
The company must therefore rely on access to debt and equity markets, asset-backed financing, customer commitments and strategic support to fund the difference.
That arrangement can work while financing remains available at acceptable prices. It becomes more difficult when the cost of capital rises at the same time that the company needs to refinance maturing obligations and continue investing in new capacity.
The balance sheet explains the 15% yield
At June 30, 2026, CoreWeave reported approximately $31 billion of recourse debt and $3.6 billion of non-recourse project debt. Operating leases add a further debt-like obligation measured in the mid-teens of billions of dollars.
Cash and restricted cash provide some liquidity, but the company’s funding requirements must be considered against its capital expenditure program and scheduled repayments. Debt maturities also create refinancing requirements over the next several years, with multibillion-dollar repayments due annually through 2028.
Annualizing second-quarter adjusted EBITDA produces gross leverage of approximately 5.9 times and net leverage of around 5 times, using the reported debt figures and adjusting for cash. On the same adjusted earnings basis, interest coverage is approximately 2.4 times.
In short, the balance sheet depends on future growth arriving as planned.
The latest trading levels show what investors are demanding in return for bearing that risk. A yield around 15% on unsecured debt reflects deep concern about the company’s ability to fund an ambitious investment program, service its debt and preserve financial flexibility under less favorable conditions.
Growth versus funding risk
The downside does not require AI demand to collapse. If hyperscalers pause to digest recent investment, contract renewals bring lower prices and older processors lose competitiveness, revenue could stall at $11bn–$13bn in 2027 and $12bn–$15bn in 2028. With margins falling into the high 30s, EBITDA could settle around $5bn–$6bn and leverage remain at 6–7 times. That would make refinancing more difficult even if the business continued to grow.
I will focus on the conversion of contracted capacity into energized megawatts, recognized revenue and cash, not headline backlog alone. CoreWeave reported $104 billion of backlog at June 30, 2026, but delivery depends on power, construction, customer demand and contract economics. Customer concentration, upcoming maturities and the terms of new asset-backed loans will also show how much financing the expansion still requires and at what cost.
The wider data-center question is whether expected AI revenues can support the capital committed to build capacity. CoreWeave shows that a large backlog and rapid revenue growth can coexist with heavy cash burn, high leverage and dependence on external financing.
The recent repricing does not, by itself, prove that the AI infrastructure boom is ending. It does show that investors are distinguishing between demand for computing power and the ability of individual projects and operators to finance, deliver and earn an adequate return on that capacity.
Alessandro Rovelli
Rovelli Research
Independent credit investment research
