Microsoft-Oracle AI Cloud Deal Winds Down: Older GPUs Go First
The GPUs Microsoft rented from a rival in 2023 because it could not get enough of them are now the ones it no longer wants. Days earlier, it asked another supplier to rebuild a site for the newest models.
Microsoft has been switching off AI computing capacity it rented from Oracle over recent months after deciding not to renew the contract, Bloomberg reported on November 6 in a story carried by Reuters. The move is part of a wider review of cloud partnerships worth billions of dollars. The rented infrastructure runs on previous-generation Nvidia chips, and it reportedly became less useful over time as Microsoft's needs changed. Oracle said demand for its OCI AI infrastructure, older GPUs included, remains strong. Microsoft did not immediately respond to Reuters' request for comment.
One expiring lease is small next to the sums both companies are spending on AI. The reason behind it is the interesting part. For most of the past three years, a GPU lease was valued by how much capacity it delivered. Microsoft is now valuing leases by which generation of silicon sits inside them, and a contract signed in 2023 is failing that test.
The 2023 deal was a product of scarcity
The arrangement dates to November 7, 2023, when Microsoft announced a multi-year agreement to move some of the machine-learning models behind Bing Search onto Oracle's GPU supercluster. The Register read it at the time as a sign that Microsoft was so short on capacity for its AI services that it would rent from a cloud competitor. Oracle then had tens of thousands of Nvidia A100 and H100 GPUs available for rent.
The new reporting does not say exactly which chips are in the capacity being shut down, only that they are a previous generation. The arc is still clear enough. Hardware that was scarce enough in 2023 to justify paying a rival is, three years later, hardware Microsoft sees no reason to keep paying for.
This is not a company pulling back on compute. When Microsoft disclosed Azure revenue in dollars for the first time on October 28, the quarterly figure was $33.2 billion, and it is still committing billions to AI infrastructure. What it is cutting is old compute, not compute in general.
The same week, IREN got the opposite instruction
That reading gets sharper next to another report from the same week. IREN, a data center operator that rents GPU capacity to large customers, said on November 5 that more than $4 billion in annual recurring revenue it had expected to start in 2026 has been pushed to January at the earliest, possibly February. It blamed a customer-requested change and supplier timing.
Co-CEO Daniel Roberts named Microsoft as the customer. "They made a request to reconfigure and upgrade some of the network infrastructure," he said, so that the 200 megawatts of capacity could serve the latest generation of frontier models. CNBC also reported that Microsoft is weighing a reduction in the number of neoclouds it counts as partners.
Put the two stories side by side. At Oracle, older capacity is being turned off. At IREN, capacity that has not even been handed over yet is being rebuilt to a higher specification. My read is that both decisions follow the same rule: what chips are inside, and whether the network around them can handle the next model. Microsoft sets that rule. Its suppliers carry the cost and the delay of meeting it.
Roberts hinted at the pressure on the supply side when he said that "capital markets have become more selective in the last few weeks." A supplier that has to rebuild a site before getting paid needs patient lenders.
Oracle is expanding on borrowed money
The rule matters most for Oracle, which has spent the autumn managing the risks of a very large buildout.
On September 24 it sent a force majeure notice to the developer of Project Jupiter, a Stargate data center campus in New Mexico built with a Blue Owl Capital unit. The notice would let Oracle delay payments if the facility misses its 2028 target to come online, although Oracle said the project remains on schedule. On November 2 it filed with the SEC to sell $10 billion in common stock and $10 billion in senior notes. The raise had been previewed at Oracle's investor day the week before, and the shares still fell 8.8% that day.
The balance sheet leaves little room for surprises. S&P cut Oracle to BBB- in July, one notch above high yield, and the company carried $125.34 billion in debt as of August 31.
Oracle's own answer deserves weight. If buyers for older GPUs are as plentiful as it says, the space Microsoft vacates could be re-let quickly, and this becomes a footnote about one contract ending. The less comfortable scenario is a pattern: large customers skipping renewals each time a chip generation turns over. In that case the new capacity Oracle is now financing with debt would face the same question a few years from now. That is a possibility, not a forecast, but this is the first time it has shown up attached to a named contract.
What would settle it
The first signal is whether Oracle names who takes over the capacity, and on what terms. A named tenant at a reasonable price would turn "demand remains strong" from a statement into evidence.
The second is how far Microsoft's partnership review goes. If the talk of fewer neocloud partners turns into ended contracts, the companies that rent GPUs for a living may have to write shorter leases or build upgrade costs into them. Either way, the economics of AI cloud capacity would start to look less like real estate and more like hardware with a short shelf life.
Sources
Comments
Post a Comment
Please read our comment policy before commenting.