Is the AI boom over? Follow the debt, then the machines

The AI building boom has acquired an awkward companion: a much closer examination of the bill.

Wednesday’s selling caught companies across the infrastructure chain. In Reuters’ intraday report on 14 October, Nebius was down 15.1%, CoreWeave 9.5%, IREN 9.9% and Applied Digital 8.8%. These were snapshots during trading, rather than closing prices.

“Investors are questioning whether the AI buildout will generate the returns that were expected,” Kim Forrest, chief investment officer at Bokeh Capital Partners, told Reuters. [Source 1]

That question has teeth. Building a computing business means paying for machines, power and buildings before all the corresponding revenue arrives. If financing becomes more expensive, the customers can remain enthusiastic while the economics for the infrastructure owner become less comfortable.

The distinction is visible in the companies’ own accounts.

CoreWeave’s interest bill is already substantial

CoreWeave reported second-quarter revenue of $2.575bn, against $1.212bn a year earlier. Chief executive Michael Intrator said: “Customer demand is accelerating.”

The same quarter contained $640m of net interest expense and a $626m net loss. Net interest expense was equivalent to approximately a quarter of revenue.

Its adjusted EBITDA of $1.510bn gives a different view of operating performance, but that measure excludes interest and other substantial costs. It cannot settle the question of what remains for shareholders after the infrastructure has been financed.

CoreWeave also reported approximately $104bn of revenue backlog at the end of June. The company’s definition makes delivery and service availability conditions explicit. That is valuable contracted business, but the backlog has to be converted into an operating service before the associated revenue can be recognised. [Source 2]

The scheduling problem underneath those figures is familiar to anyone who has commissioned complex equipment. Spending begins before acceptance. Delays can stretch the interval between paying for an asset and earning from it.

Financial leverage makes that interval more consequential. As a simple illustration, an additional three percentage points of annual interest on $1bn of borrowing represents $30m a year. That is arithmetic, not a forecast for CoreWeave, but it explains why a change in the price of money can alter the value of a project without changing the number of customers asking for compute.

IREN shows why the contract matters

IREN’s transition also needs to be read through actual revenue and financing terms.

For the financial year ended June 2026, it reported $128.8m of AI cloud revenue and $578.2m from bitcoin mining. The historical revenue mix therefore still reflected the business it was moving away from.

Its financing disclosures show different economics within the new business. IREN reported $3.6bn of investment-grade GPU financing associated with its Microsoft contract at a weighted average interest rate of 6%, excluding fees.

A separate $2.4bn financing led by Blue Owl and PIMCO for its Mackenzie air-cooled expansion carried a fixed rate of 9%. That deployment served non-investment-grade customers.

The facilities have different structures and counterparties, so the rates are not interchangeable measures of IREN’s overall borrowing cost. They do demonstrate that the customer behind the GPU deployment matters to its financing. [Source 3]

There is a second distinction inside the growth figures. Contracted annualised run-rate revenue describes a level of business under specified operating assumptions. It is not the revenue already earned during a financial year. Commissioning, testing and customer acceptance sit between the contract and the realised result.

A market can become less willing to pay today for that future conversion without concluding that the contracted workload has disappeared.

The equipment supplier is telling another part of the story

ASML’s latest results make a blanket declaration that the AI boom is over difficult to sustain.

The lithography equipment maker reported third-quarter sales of €11.7bn and net income of €5bn. More significantly for the capacity argument, it said its planned supply for 2027 and 2028 would not fully meet customer demand.

Chief executive Christophe Fouquet said that “some customers are expressing an interest in placing orders with us for 2029 and even 2030”.

ASML is investigating additional 2028 capacity beyond the increase it had previously communicated. Customer interest extending into later years is not the same as completed orders, but the company is describing pressure to supply more equipment, rather than a retreat from investment. [Source 4]

The businesses also occupy different positions in the spending chain. A chipmaker buying a lithography tool, a cloud provider financing GPUs and a server manufacturer supplying an installed system do not earn the same margin or take the same risk. Treating them as a single “AI trade” obscures those differences.

Dell and HPE were caught in Wednesday’s sell-off too. Reuters reported falls of 10.3% and 6.4%, respectively, alongside Barclays downgrades. The bank said it did not yet see a meaningful AI inflection in their businesses. [Source 1]

More hardware revenue is only part of the financial story. The margin attached to that revenue, the capital needed to support it and the time taken to collect the cash determine how much value reaches the company funding the expansion.

The evidence so far supports a harder examination of returns. It does not establish that demand for AI hardware has collapsed. CoreWeave’s interest expense shows the burden of financing the expansion already under way; ASML’s capacity outlook shows customers still asking the manufacturing chain to expand further.

Both can be true for a considerable time. The difficult part of this boom is turning expensive, functioning infrastructure into enough cash to justify what was paid to build it.