CoreWeave, debt and data centres: the AI bubble is first a question of financing
· Tarek Nachnouchi
In mid-October, the bonds of CoreWeave, a company that rents out computing power for AI, reached crisis-level yields. The analysts quoted by CNBC consider demand real and worry about how the data centres are financed. My thesis: the price and availability of computing power will stay volatile, and a European company should avoid depending on any single supplier.
What happened
On 15 October, according to CNBC, the yield on a CoreWeave bond maturing in 2031 reached 15.78%. The same day, according to Bloomberg, the cost of insuring its debt against default for five years hit a record 1,194.5 basis points during the session, which the market reads as a more than 60% chance of default within five years. An analyst at S&P Global Market Intelligence notes that from 1,000 points, the market generally regards a borrower as distressed. Evercore ISI's trading desk, quoted by CNBC, sums up the mood: the market now refuses AI debt that has no operating cash flow to support it. Over the week, the share price fell by about 25%, to $61 on 16 October. According to Reuters, bonds issued in June at 9.675% were then yielding 15%.
Reuters explained the model on 16 October. The sell-off followed quarterly results from large cloud providers that hinted at data centre delays and slower investment growth. CoreWeave borrows to buy chips, land and energy, then rents the capacity to customers such as Microsoft, Meta or OpenAI. The rent comes later. In 2025, the company reported $5.1 billion in revenue and a net loss of $1.2 billion. Its 2026 capital spending is estimated at $35 to $39 billion. According to Axios, its free cash flow was minus $5.7 billion in the latest reported quarter.
The company has arguments. Its order backlog exceeded $104 billion at the end of June, and Nvidia has agreed to buy its unsold capacity until April 2032. Its CEO wrote to staff that markets can "misunderstand and act irrationally". Gartner, quoted by Reuters, nevertheless expects 80% of "neoclouds" to have lost their independence by 2030.
Why I talk about financing
Axios puts the shift well. The data centre boom was first paid for by the profits of Amazon, Meta and Alphabet. It now depends on money raised in the bond markets. Bloomberg reported on 17 October that high-yield bond buyers are shutting out tech companies looking to borrow.
The strain also reaches more solid players. According to CNBC, the cost of insuring Oracle's debt against default rose by about 10% in a week. I read this as the end of easy money for building AI.
Étrechet, or the timeline seen from the village
On 14 September, in Étrechet, a village of 1,000 people near Châteauroux, four municipal councillors, including the three deputy mayors, resigned. The mayor, Florence Laurent, followed the next day. The cause: Google's data centre project on the Ozans site, about 200 hectares. According to France 3, the mayor cites visual nuisance, PFAS, house values and road traffic. A former deputy mayor told franceinfo that since planning powers moved to the wider urban authority, the village has one vote out of 53. On 17 October, about 800 people demonstrated on the site. Construction is planned from 2027, and a public debate begins in November.
Nationally, according to RTE, the French grid operator, quoted by The Local, nearly 18 gigawatts of grid capacity had been reserved by May 2026 for around 80 data centre projects. In Bordeaux, the city scrapped the planned site for a 250-megawatt AI campus.
For an executive, the CoreWeave story and the Étrechet story meet. The computing capacity you are promised depends on financing and permits that can slip.
Three scenarios over 18 months
| Scenario | What happens | Effect for a European company | Indicator to watch |
|---|---|---|---|
| 1. Orderly landing | Large customers honour their contracts, neoclouds refinance at a higher cost | Stable computing prices, a few targeted increases | CoreWeave bond yields below 10% |
| 2. Credit squeeze | The most indebted projects are delayed or cancelled | Delays in accessing the heaviest models, price rises | Data centre delays announced by hyperscalers |
| 3. Consolidation | Fragile neoclouds are bought by stronger players | Fewer suppliers, tougher contract terms | Announced acquisitions and backstops |
I think scenario 3 is the most likely over 18 months, with episodes of scenario 2. This is a personal conviction, with no numerical model behind it.
What I recommend
- Map your dependency. For each AI use, note which model you use, with which supplier and on which infrastructure. Ask your suppliers where the computing they resell to you actually runs.
- Keep a tested alternative. For critical uses, validate a second model or a second supplier, and check that it works with your data.
- Read the pricing clauses. Spot the clauses that let the supplier change prices or quotas during the contract.
- Prefer short commitments. While computing prices remain volatile, I would limit volume commitments to 12 months.
- Track three indicators. Neocloud bond yields, data centre delays and price increases from the major model providers.
I map this dependency with every client before deploying AI solutions.
My takeaway
AI's problem lies in the gap between spending, which is immediate, and revenue, which comes later. For a European SME, the best protection remains the ability to change supplier without rebuilding everything. I will review these three scenarios in three months.
If you would like to discuss this for your organisation, I offer a first conversation of 30 to 45 minutes, free and with no commitment.