The pressure building around the AI investment cycle is no longer just a valuation argument. It is increasingly a capital-cost and capital-structure argument.
Compute demand remains exceptionally strong. The harder question is who earns a return above the cost of financing the servers, data centers and long-duration contracts required to meet that demand.
The neocloud problem is return on capital, not near-term demand
On Wednesday’s Bloomberg Surveillance, DA Davidson analyst Gil Luria framed the issue starkly: some neocloud businesses are generating operating returns far below the cost at which they borrow. He cited one example at roughly a 1% return against borrowing costs around 9%, and argued that many neoclouds may not exist in their current form five years from now.
That is an analyst judgment, not an established outcome. But the structure of the argument matters. If capital markets eventually refuse to refinance a weak operator, the GPUs and data-center assets do not disappear. They can be sold, refinanced or absorbed by hyperscalers and better-capitalized infrastructure owners.
AI demand can survive an equity wipeout
Luria’s most useful distinction is that neocloud failure does not necessarily equal AI-demand failure. Microsoft, Amazon, Google, Meta and other large buyers can build more capacity themselves, while temporarily using neoclouds to absorb spillover demand under shorter contracts.
That implies a scenario in which AI hardware demand stays strong even as some equity holders and subordinated creditors lose money. The economic value migrates toward the owners with cheaper capital, durable customer contracts and the ability to finance replacement capacity.
AI concentration is spreading through multiple asset classes
Carlyle’s Akhil Bansal made the complementary point on Bloomberg: AI exposure now reaches investment-grade bonds, commercial real estate, private credit and public markets. Investors therefore need to trace the underlying financing structure rather than assume that nominally different securities provide diversification.
The key questions are increasingly granular: is the investor taking hyperscaler counterparty risk, residual GPU-value risk, construction risk, power-price risk or refinancing risk? Those exposures can behave very differently even when every asset carries an “AI infrastructure” label.
What to watch next
| Signal | Why it matters |
|---|---|
| CoreWeave debt yields and CDS | Shows whether the marginal cost of AI infrastructure capital is still rising. |
| OpenAI and Anthropic fundraising terms | Tests whether the largest model companies can continue raising capital at extraordinary scale. |
| Hyperscaler and SpaceX capex | Shows whether the deepest-pocketed buyers keep expanding through financing stress. |
| Neocloud IPOs and refinancing | Measures public-equity and credit appetite for the most capital-intensive structures. |
| Data-center permitting and power obligations | Political restrictions can extend construction timelines and raise the financing hurdle. |
Market takeaway: the market does not need AI spending to collapse for financing stress to matter. A move from “everything AI can get funded” to “only the strongest balance sheets can fund cheaply” is enough to change winners, losers, recovery values and multiples across the complex.