The AI build-out has a new problem: the chips it bought last year are already too old

For two years the question in AI infrastructure was whether enough computing could be built. The question now is who owns the machines when they stop being the right machines.

Microsoft has been deactivating rented server capacity at Oracle and will not renew an agreement signed in 2023, according to Bloomberg, which reported the move this week. The Oracle-hosted systems ran older generations of Nvidia processors that no longer match what Microsoft's workloads need. Neither company disclosed the value of the original contract, which supported AI features in Bing.

The same reporting describes a broader review: Microsoft is reassessing more than $80 billion of outside cloud commitments, including deals with CoreWeave, Nscale and Nebius, after costs and delivery delays accumulated. It recently declined an option to more than double a Texas deployment with Nscale — an extension that would have been worth tens of billions — weeks after pulling out of Nscale's planned site in Loughton, England, where power connections had slipped toward the 2030s.

$80bn

of outside cloud commitments under review at Microsoft

notch above junk: Oracle’s credit rating

2.5 yrs

commercial life of one accelerator generation, against 5–6 years on the books

Demand did not fall. The hardware aged

The distinction matters, because both things can be true at once.

AI accelerators are not generic servers. Each generation improves performance per watt, and a data centre is constrained by power and cooling long before it runs out of floor space. A hall full of three-year-old cards can be simultaneously fully booked and economically obsolete: it costs the same to power, delivers fewer tokens per megawatt, and cannot serve the largest models at the latencies customers now expect.

Accounting life5.5 years

what the depreciation schedule assumes

Commercial life2.5 years

how long the generation stays competitive at the frontier

The gap between the two is where the write-down sits.

That is why the market for "older-generation hardware" exists but thins quickly. Oracle says demand for its AI infrastructure remains robust, including for earlier-generation equipment. Landlords rarely describe their own inventory as stranded.

The landlord carries the risk

The shift lands hardest on companies that borrowed to build capacity and then rented it out.

A hyperscaler rents the capacity
The asset stays off its own balance sheet, so obsolescence sits with the landlord.
The hardware ages
Same power draw, fewer tokens per megawatt, and the largest models no longer fit the latency budget.
The contract stops matching the machine
The paper still says years; the equipment is worth less than the paper.
The landlord books it
A write-down is not a demand collapse. It is the balance sheet admitting the new value.

The chain runs from the renter to the owner of the metal.

Oracle has spent heavily on AI data centres, and the spending has pushed free cash flow negative. Its credit rating sits one notch above junk, and the stock has more than halved since a June peak. Shares fell about 3% on the Microsoft report. An earlier attempt to lease additional capacity to Microsoft reportedly collapsed in June.

The neoclouds have the same shape of exposure with less balance sheet: CoreWeave, Nscale and Nebius all built their businesses on long contracts with a handful of large customers. Those contracts look like revenue to the seller and like a liability to the buyer when the hardware underneath them ages faster than the paper.

The accounting arrives later

Gartner has warned that aggressive capital spending on AI infrastructure risks significant write-downs as faster and more efficient chips reach the market. That is the quiet part of the cycle. A write-down is not a demand collapse; it is the moment a company admits the asset is worth less than the balance sheet says.

Depreciation schedules assume useful lives of five to six years for this equipment. The commercial life of a generation of accelerators is turning out to be shorter — perhaps two to three years at the frontier, longer at the edge.

What changes now

Three things follow, and none of them is "AI is over".

First, contracts get shorter and exit terms get more attention than headline value. A ten-year commitment at a fixed price is only attractive if the hardware inside it stays competitive, and buyers have learned that it may not.

Second, the marginal buyer becomes more careful about who owns the residual value. Renting shifts obsolescence to the landlord, which is why the hyperscalers are renting — and why the renters' credit is now part of the story.

Third, the industry's growth numbers and its asset quality start to diverge. Revenue can keep rising while returns on the capital already spent fall, and the gap is where the next round of bad news usually sits.

The build-out is not slowing because nobody wants the compute. It is slowing where the compute is oldest, and someone has to book the difference.

Source: Bloomberg and Financial Times reporting on Microsoft's cloud commitments, as summarised by Times of India; Gartner guidance on AI infrastructure spending; Oracle's published financials.

Contact: [email protected]