Microsoft Ends Oracle AI Rental Deal Amid $80bn Commitments

Daniel Okafor

Microsoft has committed more than $80 billion to rent AI capacity from CoreWeave, Nscale and Nebius, even as it ends an Oracle agreement announced in 2023.

The reversal forms part of Microsoft’s reassessment of cloud partnerships running into billions of dollars, while the company continues to spend heavily on capacity from specialist providers.

Bloomberg reported that Microsoft has been shutting off Oracle capacity in recent months after deciding not to renew the contract. People familiar with the matter said the rented infrastructure no longer met Microsoft’s specifications and used an older generation of Nvidia chips.

In the AI cloud infrastructure market, the decision separates Microsoft’s technical requirements from Oracle’s broader claim on demand for older accelerators. It does not mean Microsoft has abandoned rented compute, because it remains committed to more than $80 billion of capacity elsewhere.

What changed in the Oracle agreement?

Microsoft did not renew the contract for technical reasons, according to the sources cited by Bloomberg. The company judged the rented capacity less useful because it no longer met its own specifications. The hardware came from an older generation of Nvidia chips, and the report did not identify specific models, server quantities or performance thresholds.

Oracle presents a different commercial picture. An Oracle spokesperson told Bloomberg: “Demand for its AI infrastructure remains strong, including for older GPUs.” On its September earnings call, Oracle said: “GPU capacity up for renewal in fiscal Q1 was renewed or resold at a 20% premium to prior contracts.” Most of those chips were four years old or older.

Both accounts can hold at once. Microsoft can reject older capacity for its own AI-assisted Bing workloads while Oracle can secure higher prices from other customers for much of the capacity it put up for renewal.

Bloomberg attributed the 20% figure to Oracle’s earnings call, rather than an independent price benchmark. Still, the commercial facts can coexist: older GPUs can retain market value even when Microsoft determines that they no longer suit its own workloads.

For operators, the lesson is specific: capacity does not have uniform commercial value. Buyers must distinguish between hardware that remains sellable in the AI cloud market and capacity that still meets a customer’s workload requirements at an acceptable operating cost and delivery schedule.

Why did the 2023 deal matter?

Oracle and Microsoft announced the agreement in 2023, with Oracle selling Microsoft access to servers to power an AI-assisted version of Bing. Bloomberg described Microsoft’s willingness to rent from a major rival as an early sign of the shortage of processing power required for AI.

The reversal captures a basic procurement change: scarcity once made external capacity strategically attractive, but workload requirements now determine whether rented infrastructure remains useful regardless of the provider’s brand or the age of its fleet. The deal therefore began as evidence of AI capacity shortages and now illustrates the risk of locking rental arrangements to hardware that a customer later considers unsuitable.

The episode also shows that a rental agreement can preserve strategic flexibility at the start and become a technical mismatch later. In 2023, Microsoft accepted a rival’s infrastructure in response to a processing-power shortage. Now, the customer treats the agreement as unsuitable and switches the capacity off, reducing its exposure to Oracle without changing its broader reliance on external compute.

Where is Microsoft placing its capacity commitments?

Microsoft has signed contracts with the neocloud providers CoreWeave, Nscale and Nebius, committing more than $80 billion in total to rent capacity from other cloud companies. That figure establishes the scale of Microsoft’s continued use of external capacity, although Bloomberg’s report does not say how much of the sum directly replaces the lapsed Oracle arrangement.

The spending has not removed Microsoft’s operational concerns. Bloomberg previously reported that senior Microsoft leaders had become frustrated with neocloud providers over high costs and delays. Around the time it allowed the Oracle agreement to lapse, Microsoft tightened oversight of other data-centre projects, pulled back from projects it regarded as risky and debated whether to sign further neocloud deals.

Taken together, those steps suggest that Microsoft still values outsourced AI capacity but will not treat a signed commitment as proof of acceptable delivery. The company’s challenge is to turn contracts with CoreWeave, Nscale and Nebius into reliable capacity at a cost its own AI services can support, while reducing exposure to projects that slip or fail internal risk reviews.

What should operators watch?

The first test is delivery under the CoreWeave, Nscale and Nebius contracts. Microsoft has already tightened control over the projects supporting its external capacity strategy, and Bloomberg has reported delays and cost complaints among senior leaders.

Providers that want to retain large customers will need to make schedules, technical fit and pricing more predictable than the experience that prompted Microsoft’s new scrutiny. The second test is the treatment of older accelerators. Oracle’s 20% renewal or resale premium supports the commercial case for four-year-old and older GPUs, while Microsoft’s decision shows why that demand does not guarantee suitability for every AI workload.

That leaves cloud buyers with a segmentation challenge: model, performance and economics matter more than treating GPUs as interchangeable inventory. No new launch date or spending timetable accompanied the report. The next observable milestones are therefore execution under the CoreWeave, Nscale and Nebius contracts and Microsoft’s decisions on additional neocloud projects, with each new agreement likely to face tighter project oversight after the Oracle deal lapsed.