The Hidden Risk in AI Infrastructure: What CoreWeave Shows Fund Investors

The most insidious risk in artificial intelligence is not a model that goes wrong. It is a financing chain that looks diversified from the outside and turns out to rest on a small number of counterparties, a single supplier and collateral that loses value on a technology timetable. CoreWeave, a listed specialist AI cloud provider, is a useful case study precisely because it discloses all of this in its public filings. Nothing in this article suggests any impropriety. The point is that the structure is visible to anyone who reads the filings, and many portfolios that never chose to own it may hold the risk anyway.

"AI risk rarely arrives as a single position. It arrives as a loan here, a convertible there, a counterparty's revenue line and an index weight, all tied to the same few balance sheets. A fund board that cannot see that through the portfolio is not managing the risk it thinks it is."David Lloyd, Chief Executive Officer at CV5 Capital

The Business in One Paragraph

CoreWeave rents graphics processing unit (GPU) capacity to companies training and running AI models. Its revenue for the quarter ended 30 June 2026 was $2,575 million, more than double the $1,212 million of a year earlier, and it reported a revenue backlog of approximately $104 billion. To deliver that backlog it is building data centres at a pace financed largely with debt: property and equipment purchases were $6,422 million in that quarter alone. For the AI build-out, that combination is the opportunity. For investors, it is where the risk sits.

Where the Risk Hides

1. The supplier is also the shareholder and the buyer of last resort

CoreWeave's own filings describe a relationship with NVIDIA that runs in several directions at once:

Each arrangement is commercially rational and disclosed. Together they mean that part of the demand for the GPUs, part of the equity funding the purchases and part of the guarantee behind the capacity come from the same company. That does not make the revenue false. It does make it harder to tell how much of the demand is independent, and it concentrates the outcome on one counterparty's continued willingness to support the arrangement.

2. Collateral that depreciates on a technology clock

CoreWeave finances much of its build through delayed draw term loans that its annual report describes as collateralised by contractual cash flows and infrastructure assets, and through equipment financing secured on the equipment itself. It depreciates its computing equipment over six years, an estimate it extended from five years with effect from January 2023, which reduced expenses in that year.

A six-year life may prove right. It is still a judgement, and it matters twice: it drives reported earnings, and it underpins the value lenders attach to the GPUs securing their loans. The same filings say the company expects to be among the first cloud providers to deploy NVIDIA's next platform. If newer chips make older ones less valuable faster than the depreciation schedule assumes, the gap shows up first in collateral coverage and resale values, and only later in reported figures.

3. Backlog is not cash

A backlog of approximately $104 billion is a large number, and more than 98% of 2025 revenue came from committed contracts. But several of the largest commitments are expressed as amounts customers have committed to pay "up to", over periods running into the 2030s, and each depends on the customer's continued ability and willingness to perform. CoreWeave's filings list OpenAI commitments of up to approximately $11.9 billion through October 2030 and up to approximately $6.5 billion through May 2031, and a Meta commitment of up to approximately $21.0 billion through 2032. The credit quality of the backlog is the credit quality of those customers over those periods.

4. Concentration is improving, but it is still concentration

Microsoft accounted for approximately 67% of CoreWeave's revenue in 2025. By the quarter ended 30 June 2026, the top three customers accounted for approximately 36%, 26% and 10% of revenue respectively. That is real diversification over a short period. It still leaves roughly 72% of quarterly revenue with three customers, and the company itself expects concentration to continue given the length of its contracts.

5. Growth can disguise leverage

Revenue more than doubled, and so did the net loss. Net interest expense was equivalent to about a quarter of revenue in the quarter. At 30 June 2026, balance sheet debt was approximately $35.1 billion, recourse and non-recourse combined, before approximately $16.3 billion of operating lease liabilities, against total assets of $77.1 billion. None of this is hidden, but headlines about revenue growth tend to arrive without it.

Figures are taken from CoreWeave's Form 10-Q for the quarter ended 30 June 2026 and its second quarter results release; the debt total is our sum of the reported current and non-current recourse and non-recourse debt lines.

How a Fund Ends Up Holding This Risk Without Choosing It

This is the insidious part. A fund does not need to own CoreWeave shares to be exposed to the same chain:

• Private credit. GPU-backed and contract-backed loans to AI infrastructure borrowers sit in direct lending and asset-based finance portfolios, often marked by models rather than markets. See evergreen private credit funds: liquidity and valuation.
• Convertibles and high yield. AI infrastructure issuers have been active in unsecured and convertible markets.
• Equity and index weight. AI suppliers, hyperscalers and their customers account for a large share of major equity indices, so a "diversified" long book can carry correlated exposure. See the AI capex trade.
• Counterparties. A fund's prime broker, lender or bank may itself be lending into the build-out. See prime broker counterparty risk.
• Strategy crowding. Many quantitative and fundamental strategies are positioned around the same AI theme, which raises the risk of correlated exits; see our analysis of the July 2026 leverage event.

Individually, each holding can pass its own risk review. Collectively, they can amount to a single concentrated bet on a small number of companies continuing to buy and finance GPUs at the current pace.

What Boards and Allocators Should Ask

What This Is Not

This is not a forecast for CoreWeave or a view on its securities. The company is growing quickly, has diversified its customer base materially within a year and has access to deep capital markets. The lesson is structural: in any capital-intensive technology build-out, the same few balance sheets can sit on every side of the transaction, and the risk accumulates quietly in places that individual position reviews do not join up. For how managers are using AI inside the fund itself, see AI hedge funds: research, risk and fund infrastructure.

Conclusion

The insidious risk in AI is that nobody has to take it deliberately. It arrives through financing structures, collateral assumptions, counterparty chains and index weights, each reasonable on its own. CoreWeave's filings show the pattern clearly because the company discloses it. Funds that read those disclosures, look through their own holdings and put the answers in front of their boards will be better placed than those that discover their concentration when the cycle turns.

Frequently Asked Questions

What is circular financing in AI?

Arrangements in which the same company acts in more than one role in a transaction chain, for example as supplier, investor and customer. In CoreWeave's case, NVIDIA supplies its GPUs, holds its shares and is obligated to buy residual unsold capacity under a September 2025 agreement.

Why does GPU depreciation matter to fund investors?

Because GPUs secure a large share of AI infrastructure lending. If they lose value faster than the depreciation schedule assumes, collateral coverage and model-based loan valuations can be overstated before reported earnings show it.

How concentrated is CoreWeave's revenue?

Microsoft accounted for approximately 67% of 2025 revenue. In the quarter ended 30 June 2026, the top three customers accounted for approximately 36%, 26% and 10% of revenue.

Can a fund be exposed without holding AI stocks?

Yes. Exposure can come through private credit, convertibles, index weights, counterparties and crowded strategies. A look-through review across all of them is the only way to see the total.

Is this a recommendation on CoreWeave?

No. The article uses CoreWeave's public disclosures to illustrate a structural risk pattern. It is not a view on the company or its securities.

Sources: CoreWeave, Inc. Form 8-K filed 15 September 2025; Form 10-K for the year ended 31 December 2025; Form 10-Q for the quarter ended 30 June 2026; second quarter 2026 results release.