Rows of AI server racks in a dimly lit data center hall, with a single technician walking the aisle past racks whose status lights have gone dark
On Friday, 6 November 2026, Bloomberg News reported that Microsoft is winding down an arrangement to rent AI computing capacity from Oracle. The deal dates back to 2023. According to the report, Microsoft has spent recent months turning off rented Oracle capacity after declining to renew the contract, and the servers involved ran an older generation of Nvidia chips.
On its own, that could pass for a routine contract expiry. It did not land on its own. The Oracle story arrived one day after Bloomberg reported that Microsoft was tightening oversight of data center projects run by the neocloud providers Nebius, IREN and Nscale, and just days after Microsoft passed on a major expansion option at an Nscale site in Texas. Taken together, the reports describe a company that is rethinking how it buys AI computing power, at a moment when it is still spending tens of billions of dollars on it.
This article walks through what has been reported, who said what, which numbers are firm and which are not, and why the Oracle piece matters for the wider market for AI cloud capacity.
Access without medium partner: Microsoft moving from Oracle AI Data Centers
What Bloomberg reported about Microsoft and Oracle
The core facts come from Bloomberg, with Reuters summarizing the report for its own subscribers. Microsoft is winding down an arrangement to rent AI computing from Oracle Corp., and the move is described as part of a larger reassessment of billions of dollars in cloud partnerships. Bloomberg’s sources, described as people familiar with the matter, said Microsoft has been switching off the rented capacity over recent months after declining to renew.
The reason given in the reporting is practical rather than dramatic. The rented capacity became less valuable to Microsoft over time because it no longer matched the company’s evolving needs. The Oracle hardware in question used an older generation of Nvidia chips.
Two details are worth noting because they are missing. Financial terms were not disclosed, and Bloomberg could not determine the full size of the deal. That matters for anyone trying to put a dollar figure on the story. Some secondary outlets have attached numbers to the arrangement, but those figures do not appear in the Bloomberg or Reuters accounts, so this article does not use them.
The 2023 deal that started it
The arrangement goes back to November 2023, when Oracle announced a multiyear agreement with Microsoft. Under it, Microsoft would use Oracle Cloud Infrastructure, known as OCI, alongside its own Azure platform to provide AI inference for Bing conversational search. Inference is the stage where a trained model answers real requests, such as analyzing search results and generating a reply. It is the part of AI that runs every time a user asks a question, and it needs a steady supply of GPUs.
At the time, the deal was a notable example of two large rivals in cloud computing cooperating because demand for AI capacity exceeded what either could build quickly. Microsoft needed chips and data center space fast. Oracle had capacity and wanted a marquee customer.
Nearly three years later, the picture has changed. The deal Bloomberg describes is not a failed partnership in the usual sense. Microsoft and Oracle have announced other collaborations involving AI computing beyond the Bing arrangement, and Oracle’s own statement frames the wind down as ordinary capacity turnover. What has ended, or is ending, is one specific rental of older GPU capacity.
Business Insider separately reported that talks for Microsoft to lease additional Oracle capacity collapsed in June. That report is secondhand here, but it fits the pattern of Microsoft being more selective about what it signs.
Why older Nvidia chips are at the center of the story
The phrase that drew the most attention in the reporting is the one about older Nvidia chips. GPUs do not stop working when a newer generation ships. A server running a 2023 era accelerator can still run models. The question is whether it can run the models Microsoft now wants to run, at a cost that makes sense.
The reporting suggests Microsoft concluded it could not. Newer AI models and heavier workloads tend to favor newer hardware, with more memory, faster interconnects and better efficiency per unit of power. Capacity that was ideal for Bing inference in 2023 can look less attractive when the workload mix has moved on.
One analysis of the news pointed out an important limit on that reading. Microsoft’s changing specifications do not necessarily mean older Nvidia chips have become obsolete across the entire market. A buyer with different needs may still find that hardware useful. That is exactly the argument Oracle made.
This is the real economic question hanging over AI infrastructure: how fast does a GPU lose its commercial value? Companies that borrow money to buy chips, and lenders who finance them, tend to assume a useful life of several years. If a top customer walks away from capacity after roughly three years, it raises a fair question about whether that assumption holds. Nothing in the reporting proves it does or does not. It only shows the question is being asked in public.
Oracle’s response: demand is strong, and prices went up
Oracle did not leave the report unanswered. An Oracle spokesperson told Reuters that demand for OCI’s AI infrastructure remains strong, including for earlier generation GPUs. The company also responded on X, saying capacity coming up for renewal had found buyers at higher prices.
Coverage of that response adds a specific number. According to the reporting, capacity that came up for renewal in Oracle’s fiscal first quarter was renewed or resold at prices roughly 20% above previous long term contracts. If accurate, that cuts against the idea that older chips are becoming hard to place.
There is no contradiction between the two positions, because they answer different questions. Microsoft’s reported view is that this capacity no longer fits Microsoft’s needs. Oracle’s view is that other customers still want it, and will pay more. Both can be true at once. The open issue is how much capacity is involved, how quickly it can be resold, and at what margin, and none of that is public.
Oracle also carries a large order book that investors watch closely. One outlet put Oracle’s remaining performance obligations, a measure of contracted future revenue, at $664 billion. That figure comes from a secondary source and should be checked against Oracle’s own filings before anyone relies on it, but the underlying point stands. A company with that much contracted work depends on customers continuing to sign, renew and pay.
This was not one headline, it was a week
The Oracle report makes more sense when placed in sequence.
On Thursday, 5 November, Bloomberg reported, in a story by Brody Ford under the headline “Microsoft Sours on Neoclouds IREN, Nebius, Nscale on Delays, Costs,” that Microsoft was tightening oversight of some data center projects run by those three providers. The report said Microsoft was pulling back from projects it considers risky and debating whether to sign any new neocloud contracts at all.
On the same day, IREN said it had delayed some infrastructure deliveries after Microsoft asked for changes to its network design.
On Friday, 6 November, the Oracle story followed.
Earlier in the week, the Financial Times reported that Microsoft would not renew a contract at an Nscale site in Texas, and CNBC added that Microsoft is shrinking its roster of neocloud partners while concentrating spending on a smaller group of suppliers and building more of its own data centers.
For readers new to the term, a neocloud is a newer cloud company built mainly around renting out GPU capacity for AI. The best known include CoreWeave, Nebius, IREN and Nscale. They typically sign long contracts with a few large customers and borrow heavily to build data centers. That structure works well when the anchor customer keeps expanding. It becomes fragile when the customer pulls back.
Nscale: the Texas option and the English site
Nscale, a London headquartered company backed by Nvidia, has been the most visible casualty so far. Bloomberg reported that Microsoft declined an option to expand its contract at Nscale’s Texas site. One analysis put the unused option at about 700 megawatts, which would have more than doubled the Texas contract. Microsoft had also walked away from a separate Nscale site in Loughton, Essex, in England, weeks earlier.
Darren Hardman, Microsoft’s CEO for the United Kingdom and Ireland, told CNBC in October that the English decision was tied to Nscale’s access to electricity. Power availability, not chip supply, was the stated constraint.
The exposure is large. According to regulatory filings cited in coverage, Microsoft accounts for roughly 42% of Nscale’s $103 billion in contracted value. That reporting also noted that Nscale has other major customers, including Anthropic, but that those customers often lack Microsoft’s creditworthiness. Nscale submitted its IPO filing in September, so the retreat came shortly before it sought to go public.
Microsoft has not exited Nscale entirely. Coverage notes that active contracts remain in Portugal and Norway.
IREN: delays at Childress, Texas
IREN’s situation involves timing rather than cancellation. The company said it handed over capacity to Microsoft in August, but a customer requested networking upgrade pushed the capacity online to January or February instead of this year. As a result, IREN moved its target for bringing more than $4 billion in annualized revenue capacity online from year end to January, with a possible further slip to late February.
The Information reported that delays at IREN’s Childress, Texas site led Microsoft to push back planned spending of about $5 billion. It was unclear how much of that gap IREN accounts for. IREN is building the site under a five year contract worth about $9.7 billion, and its AI revenue more than doubled to $200 million.
Nebius and CoreWeave: the exposure behind the headlines
Nebius disclosed Microsoft contracts worth roughly $17.4 billion, according to one tally. A separate analysis combined the disclosed Microsoft contracts at Nscale, Nebius and IREN and arrived at as much as $70.9 billion, though that total is the analyst’s own arithmetic from SEC filings, not a company figure.
CoreWeave, the largest name in the group, was reported to have drawn roughly two thirds of its 2025 revenue from Microsoft, according to a secondary analysis. That number should be verified against CoreWeave’s filings before reuse.
Microsoft’s own numbers: still spending, but differently
It would be easy to read these headlines as Microsoft cutting back on AI. The reporting does not support that.
Microsoft reported $45.2 billion in capital expenditures and finance leases for its most recent quarter and said capex spending will accelerate in the second half of its fiscal year, according to CNBC’s account of the company’s earnings. Separately, one outlet noted that even as Microsoft retires the Bing era Oracle lease, it continues to commit more than $80 billion to outside cloud capacity, including contracts with CoreWeave, Nscale and Nebius.
So the reported change is not about whether to buy AI capacity. It concerns where, from whom, and on what terms. Reuters framed the Oracle report as adding to signs that Microsoft is rethinking how it secures computing capacity to get more out of its massive AI investments.
Analysts offered several explanations for why a company might shift in this direction. One view, published by Simply Wall St on Yahoo Finance, argued that stepping back from Oracle, Nscale, IREN and others reduces reliance on third parties whose delays, power constraints and older chips can make AI spending less predictable. It added that pulling more buildout into Microsoft’s own data centers lets management match capacity decisions more closely to its bond funded AI program and internal return targets. That is an interpretation, not a statement from Microsoft, and the company has not publicly confirmed these motives.
The same piece suggested investors watch net debt and free cash flow relative to AI infrastructure spending over the next few quarters as neocloud contracts roll off.
How the market reacted
Market moves were uneven, which makes them hard to read.
On Thursday, as the neocloud oversight story circulated, Nebius fell about 16% and IREN fell about 16%, according to one account. Benzinga later reported that CoreWeave, Nebius and IREN were each down roughly 45% over the past month, even after rallying on Friday.
Oracle’s reaction depends on who is counting. Some outlets reported Oracle shares falling around 3%, others more than 4%, with one describing a two year low. Reuters, however, described Oracle and Microsoft shares as flat in afternoon trading. These figures reflect different moments in the session, and readers should check closing prices from an exchange or brokerage rather than rely on any single article.
Microsoft’s stock showed a small decline in the quotes displayed alongside several reports, not a dramatic move. That is consistent with the idea that investors treat this as a supplier and neocloud story more than a Microsoft problem.
What is confirmed and what is not
Because this story has spread quickly through secondary outlets, it helps to separate the layers.
Confirmed by Bloomberg and Reuters: Microsoft is winding down an arrangement to rent AI computing from Oracle. The arrangement dates to 2023. Microsoft has been turning off rented capacity in recent months after declining to renew. The hardware used an older generation of Nvidia chips. Oracle says demand for its AI infrastructure remains strong, including for earlier generation GPUs. Financial terms were not disclosed.
Reported by named outlets, not independently confirmed here: The Nscale Texas option and Loughton exit (Bloomberg and the Financial Times). IREN’s delay and the roughly $5 billion spending push back (IREN’s own statement and The Information). Oracle’s claim of about 20% higher pricing on renewed capacity (Oracle statements as relayed in coverage). CNBC’s report that Microsoft is shrinking its list of neocloud partners.
Unknown: The size of the Oracle deal, how much capacity remains active, which neoclouds Microsoft will ultimately drop (CNBC reported Microsoft is still evaluating), and whether Oracle has already resold the freed capacity in full.
Neither Microsoft nor Oracle has laid out a complete public account. Much of what is known comes from anonymous sources and company statements issued in response to press reports.
Why this matters beyond Microsoft and Oracle
The immediate stakes are narrow. One contract is ending, and the company that provided it says it has buyers. The wider significance is about what the AI infrastructure boom is built on.
For neoclouds, the central risk is customer concentration. When one buyer accounts for 42% of contracted value, or roughly two thirds of annual revenue, that buyer’s decisions function as the company’s decisions. Their stock performance this month reflects how much weight investors place on that.
For chip buyers and lenders, the issue is depreciation. Financing for GPU fleets often rests on assumptions about how long hardware stays valuable. Oracle’s report of renewals at higher prices supports the optimistic view. Microsoft’s reported decision supports the cautious one. Both facts now sit in the same news cycle.
For Microsoft, the shift suggests a preference for control. Building its own data centers means more direct responsibility for power, construction timelines and cost, but fewer dependencies on partners that have, in several cases, run into delays.
For Oracle, the next few quarters will show whether its claim about demand holds up in reported results, including how the company discloses utilization and renewal pricing for older GPU capacity.
What to watch next
Several developments could clarify the picture in the coming weeks. Microsoft’s decisions on which neocloud partners to keep, which CNBC said are still under review. IREN’s progress toward its January delivery target and whether the late February scenario materializes. Nscale’s public disclosures as it moves through its IPO process. Oracle’s commentary on resale of freed capacity. And Microsoft’s next capex disclosure, which will show whether the company is shifting more spending into its own facilities.
The simplest reading of the week is that Microsoft is spending heavily on AI but is becoming choosier about who supplies the capacity. The Oracle wind down is the oldest and least dramatic piece of that shift, since it involves 2023 hardware and a Bing workload. The Nscale and IREN developments are newer and carry larger numbers. Whether they add up to a lasting change in how the biggest AI buyers purchase computing power will depend on contracts that have not yet been signed, or renewed, or dropped.