Microsoft has reportedly declined to renew an Oracle AI computing agreement and has been switching off the rented capacity in recent months. The arrangement dates to 2023, when the capacity was intended to support Microsoft’s AI-assisted Bing search service.
Bloomberg first reported the wind-down on November 6, 2026, citing people familiar with the matter. It described the decision as part of a broader reassessment of billions of dollars in cloud partnerships.
The reported mismatch between the hardware and Microsoft’s current requirements is the key detail. According to Data Center Dynamics’ November 9 follow-up, the agreement covered older-generation Nvidia hardware that no longer met Microsoft’s specifications. Newer Oracle contracts appeared unaffected.
Microsoft has not publicly confirmed the nonrenewal in the cited coverage. The agreement’s value and the amount of capacity switched off were not disclosed.
Those limits matter. The evidence points more directly to a capacity-fit decision than to falling overall AI demand. It also shows why cloud buyers and infrastructure investors need to look beyond the number of GPUs covered by a contract.
Why Older AI Capacity Can Lose Its Value
The reported Oracle arrangement began in 2023. By late 2026, Microsoft’s requirements had apparently moved beyond what that particular capacity offered.
That timeline does not identify the chips. The reporting names neither a specific Nvidia model nor the cluster’s memory, networking or configuration. It would be a mistake to infer that the GPUs were exactly three years old, or that they had become incapable of running AI software.
The narrower finding is still consequential: usable hardware can become a poor fit for a particular customer.
AI capacity is a bundle of capabilities. GPU memory affects which models and workloads fit on a device. Connections between GPUs affect how efficiently they can work together. Power consumption, software support and utilization influence the cost of completing useful work.
A newer generation can change that calculation, but the chip alone is not the whole calculation. A buyer may need a different balance of memory, interconnect capacity and deployment flexibility.
My reading is that this is the most useful distinction in the Oracle story: hardware obsolescence for one customer is different from hardware becoming worthless. A configuration that no longer suits Microsoft could still serve another workload.
Oracle makes a stronger commercial claim. It told Bloomberg that demand for its AI infrastructure remained strong, including for older hardware, according to DCD. That is Oracle’s account of its business, not an independent measurement of market demand. Still, it is consistent with different buyers valuing the same generation of equipment differently.
Networking and Power Can Change the Deal Too
Microsoft’s reported scrutiny extends beyond Oracle. DCD, citing Bloomberg, described increased oversight of projects involving specialist cloud providers Nebius, Iren and Nscale. These providers, often called neoclouds, focus heavily on supplying AI computing capacity.
Two reported examples show why a capacity contract cannot be judged by its GPU allocation alone:
- Iren: A Microsoft-requested networking infrastructure change pushed a data center project back by one to two months.
- Nscale: Microsoft withdrew from the Loughton data center project in the UK after the electricity grid proved unable to provide a connection before the 2030s.
These are different constraints. A network redesign changes how a cluster is built and connected. A delayed grid connection changes when it can operate at all.
Neither example establishes what happened inside the Oracle arrangement. They do, however, explain why Microsoft might review several infrastructure commitments at once without making one uniform judgment about AI demand.
For distributed AI workloads, networking is part of the computing system. GPUs exchanging data can spend time waiting if the connections between them cannot support the workload efficiently. Changing that design can therefore be important enough to justify a delivery delay.
Power creates a more basic dependency: promised capacity cannot serve customers until it can be energized. The issue also connects with Zeniteq’s coverage of xAI’s Colossus 2 adding eight 50-MW jet turbines, another example of energy infrastructure entering the AI buildout.
Microsoft told Bloomberg that it manages infrastructure partnerships:
“against our requirements and customer demand, including monitoring delivery and adjusting plans when circumstances change.”
DCD reproduced that statement. It describes Microsoft’s stated approach, while the individual project reports show the kinds of changes such an approach can involve.
What Cloud Buyers Should Put in the Contract
The practical lesson is to evaluate usable capacity delivered on time, not simply contracted GPU counts.
The published reporting supports three separate checks:
| Reported issue | What buyers should examine |
|---|---|
| Older Oracle hardware no longer fitting Microsoft’s needs | Hardware specifications, renewal options and opportunities to change configurations |
| Iren’s networking change delaying delivery | Agreed network design, acceptance criteria and responsibility for redesign costs |
| Nscale’s delayed grid connection | Power availability, connection milestones and remedies for late delivery |
These are implications drawn from the reporting, not disclosed terms of Microsoft’s agreements.
For a buyer, I’d put particular weight on renewal and acceptance provisions. A low headline rental price offers limited comfort if the equipment no longer fits the workload when the next contract period begins. Likewise, a delivery date is less useful if the contract leaves unclear which network configuration must be ready on that date.
Flexibility has a cost, too. A provider financing a large installation benefits from firm commitments, while a customer benefits from options to resize or change equipment. Negotiating that tension is part of the economics of rented AI capacity.
Frequently Asked Questions
4 questions
1Has Microsoft Confirmed the Oracle Deal Is Ending?
Not in the cited reporting. Bloomberg attributed the decision to unnamed people familiar with the matter, and DCD reported that Microsoft declined to comment.
