Microsoft Is Winding Down Its Oracle AI Cloud Rental: What Is Reported

Microsoft is winding down a deal to rent AI computing capacity from Oracle, according to a Bloomberg report published November 6, 2026. The report says Microsoft has been turning off the rented capacity in recent months after choosing not to renew the contract. Oracle's spokesperson says demand for its AI infrastructure remains strong. Microsoft declined to comment. Everything below is attributed to that reporting, because neither company has confirmed the reason or the terms.

Update (November 6, 2026): Reuters labels the same story

Reuters also carried the story under the headline “Microsoft is winding down Oracle deal to rent AI cloud capacity”, and its URL attributes it to Bloomberg News. We could not open the Reuters page, so we rely on that headline and URL only. We still treat the underlying claim as reported by Bloomberg, with Reuters relaying it rather than confirming it independently. We have not seen new numbers on deal size or an on-record reason from Microsoft.

TL;DR: what is reported

What developers are saying

The Hacker News discussion stayed small on November 6, with two top-level comments, but it shows how readers split on the story.

User tonoto argues that GPU rental is not a one-off capital expense because customers expect fresher generations over time, and asks how the business becomes profitable before what they call a bubble burst — a reminder that financial readers are pricing depreciation and refresh, not FLOPS alone (comment).

User neural_thing calls the story a "nothingburger," reads it as Microsoft declining to renew capacity on 2023-era A100-class hardware that powered Bing in favor of newer in-house silicon, and suggests Oracle can still re-rent that footprint on the spot market (comment).

Treat both as opinion and inference, not confirmation of chip vintages — Bloomberg's sources named "older Nvidia chips," but Microsoft has not stated models.

What Bloomberg reported

The paywalled Bloomberg article is the original report. We could not read it directly, so we rely on Investing.com's account and a MarketScreener summary, both of which cite Bloomberg. Per those accounts:

• Microsoft has been shutting off the rented Oracle capacity over recent months after declining to renew the contract.
• Citing sources, the hardware used older Nvidia chips that no longer matched Microsoft's requirements.
• The arrangement dates to 2023, when it helped Microsoft add computing power for an AI-assisted Bing.
• Bloomberg framed the move as part of a wider review of cloud partnerships, per one summary.

That last point is thinly sourced in what we read, so treat it as a framing, not a confirmed program.

What Oracle and Microsoft said

Oracle's spokesperson said demand for its AI infrastructure remains strong, including for earlier-generation GPUs. Investing.com adds that in September Oracle said it had renewed contracts for four-year-old GPUs at a 20 percent premium. That is the company's own claim and is the opposite signal from the one in the Bloomberg report: older chips still renting at a premium.

Microsoft declined to comment, so the reason for not renewing is only the sources' account.

Why this is a small data point with a big shadow

Rental of GPU capacity between rivals was one of the odd features of the 2023 to 2025 supply crunch. Microsoft, Oracle, CoreWeave and others filled each other's gaps. A non-renewal of one contract tells us little about total AI demand. But the market reads it through a nervous lens: the same week, the Nasdaq Composite fell 5.5 percent and the Philadelphia Semiconductor Index fell 12 percent on Thursday, per Investing.com, amid worries about AI infrastructure spending. We covered that selloff in AI stocks crash on November 5.

The older-chip question

The central claim is that the capacity relied on older Nvidia GPUs. Two readings fit the facts:

1. Obsolescence. Newer training and inference jobs need newer chips and more memory per GPU. Capacity on older parts loses value, and a tenant with its own newer fleet has little reason to renew.
2. Strategy. Microsoft may prefer to run capacity it controls rather than rent from a competitor, especially as its own data centers come online.

Oracle's September claim of renewed four-year-old GPU contracts at a premium argues that other tenants still value older chips. Both can be true: Microsoft's workloads may differ from the customers renewing.

What this means for people who build with AI

For builders, the practical effects are indirect:

• Prices and availability. If rental capacity frees up, spot and reserved GPU prices could soften at the margin. We have no data showing that yet.
• Concentration. Large labs increasingly depend on a few clouds. A shift in who rents from whom changes where capacity appears. See our look at OpenAI weighing a reset of 2027 data center capacity, where much of the infrastructure is built and held by partners such as Oracle.
• Siting politics. The build-out also faces local resistance. OpenAI answered with a community compact and policy blueprint and a rebrand to "AI compute campus".

If you rely on one cloud for inference, now is a reasonable time to check that your stack can move. Our guide to Nvidia and Brightspeed's planned inference network shows another way capacity is spreading outside the big data-center model.

Chips and servers in a data center, the hardware behind the Microsoft Oracle AI cloud rental story

Timeline: how the rental fits the larger picture

• 2023: Microsoft agrees to rent Oracle capacity to add computing power for an AI-assisted Bing, per Investing.com's account of the Bloomberg report.
• Recent months: Microsoft declines to renew and shuts off the rented capacity, per the report.
• September 2026: Oracle says it has renewed contracts for four-year-old GPUs at a 20 percent premium, per Investing.com.
• Thursday, November 5, 2026: A sharp tech selloff hits chipmakers and neocloud companies.
• Friday, November 6, 2026: Bloomberg publishes the report and Oracle shares fall.

The sequence matters. The contract was not cancelled on Friday; Bloomberg reports it has been running down for months. What changed on Friday was disclosure, landing on a market already on edge.

Questions readers are asking

Is Microsoft cutting AI spending? Nothing in the reporting we read says so. A non-renewal of rented capacity is consistent with Microsoft using its own newer fleet instead.

Is Oracle in trouble? The report concerns one customer and one contract on older chips. Oracle's on-record statement points the other way. The next earnings call will show whether renewals hold.

Does this affect Azure customers? The reporting does not describe any effect on Azure service. It concerns capacity Microsoft rented to supplement its own.

Why does chip age matter so much? Because the price a GPU can command falls as newer generations arrive, and because financing and depreciation assumptions depend on how long the chips stay useful. A contract not renewed for that reason draws attention from investors.

What is confirmed and what is not

What would turn this report into a trend

A single rental that is not renewed becomes a trend only if more evidence stacks up. Useful signals include other hyperscalers returning rented capacity, GPU rental prices for older parts falling across providers, and Oracle disclosing weaker renewal rates on earlier-generation systems. The opposite signals, such as more premium renewals like the one Oracle cited in September, would weaken the pessimistic reading. Readers who follow AI infrastructure should track those three data points rather than the share-price move on any one day.

How to follow this story

Watch for three things. First, Oracle's next earnings call, where analysts will ask about renewals of older-GPU capacity. Second, any Microsoft statement or filing on capacity commitments. Third, signs that other tenants are also shifting. Until then, avoid reading one contract as a verdict on AI demand.

Another pressure point is compute commitments at the other labs. Our post on Anthropic's compute outlook shows how widely stated gigawatt figures can differ between sources, a good reminder to check the numbers behind any capacity story.

Why the way this story is told matters

Capacity stories invite two lazy readings. The bearish one says a lost rental proves the boom is ending. The bullish one says a lost rental on old chips proves nothing. The honest position sits between them. A single non-renewal, on hardware that is several generations old, says little about total demand. It does say something about how fast AI hardware ages in the market, and that matters for anyone who financed servers on the assumption of long useful lives.

For a reader learning AI infrastructure, the key concept is that GPU capacity is not one commodity. Newer parts with more memory and faster interconnects serve frontier training and large inference jobs. Older parts can still serve smaller models, batch jobs and fine-tuning. A cloud with a mix can keep both busy, which is the argument in Oracle's statement. A tenant with its own new fleet, as Microsoft is building, may have no use for the older tier. If the reports hold, the same hardware can look wasted to one buyer and valuable to another.

Related reading

• AI stocks crash on November 5: what drove it
• OpenAI weighs resetting 2027 data center capacity
• OpenAI National Community Compact and policy blueprint
• OpenAI and the "AI compute campus" rebrand
• Nvidia and Brightspeed inference network
• Anthropic compute outlook: claimed vs. verified

Details are accurate as of November 6, 2026 and rest on a paywalled Bloomberg report relayed by other outlets. We will update this post if Microsoft or Oracle confirm details.