Microsoft scales back rented AI compute: what operators can take from it

On 6 November 2026 Bloomberg and CNBC reported within a short time of each other on how Microsoft buys compute. Bloomberg says Microsoft is not renewing a 2023 contract with Oracle and is switching off the rented capacity. CNBC says Microsoft wants to shrink its roster of neocloud partners. Both rest on anonymous sources, and the market chatter around them is worded more strongly than the articles.

A neocloud is a provider that rents out GPU compute for AI and is not one of the classic hyperscalers such as Amazon, Google or Microsoft. Examples are CoreWeave, Nebius, Nscale or Lambda. Many build their data centers with borrowed money and depend on a few large customers who commit capacity through multi-year contracts. If one of those customers drops out, the provider feels it directly.

What the reports actually say

I read the CNBC article in the original. The Bloomberg text is behind a paywall and could not be retrieved. For the Oracle story I therefore rely on The Next Web and MarketScale, which summarize Bloomberg. That is secondhand reporting.

Worth noting for context: according to CNBC, Microsoft has not decided to leave all neoclouds. The two sources say it wants to keep working with a few key suppliers while using more of its own data centers. The line in some newsletters that Microsoft wants to "stop renting more compute" goes beyond both reports. What is reported is one non-renewal at Oracle and an ongoing review of the neoclouds.

MarketScale adds that Bloomberg could not determine the size of the Oracle contract. Microsoft declined to comment there.

Further details in the CNBC piece

CNBC names three details that fit the picture, each coming from the parties involved. IREN said in its quarterly report that the data center capacity was handed to Microsoft in August. Microsoft then asked for a networking upgrade, so the site goes live in January or February. CFO Amy Hood said in October that some GPU contracts had slipped to the first half of 2027. And Microsoft walked away from a Nscale site in England. Microsoft's UK chief put it down to the power connection Nscale was supposed to secure.

For providers this means demand is not the only test. Capacity has to arrive on time and with the right equipment. That reading is ours, and CNBC does not phrase it that way.

The same source reports that CoreWeave said at its 2025 IPO that 62 percent of its 2024 revenue came from Microsoft. CoreWeave and Nebius report quarterly results the week from 9 November, according to CNBC. Those reports will show what the providers say about their Microsoft contracts themselves. Share price moves are not part of this piece.

What renters of GPU capacity can take from it

The reports give no rental prices, no contract terms and no availability figures. What follows is inference, not measurement.

Older cards are the likeliest to be affected at renewal. According to Bloomberg, the Oracle case involves capacity on an older chip generation that no longer fit the requirements. Anyone renting a generation that a large customer is just leaving should ask about the contract end and about what happens to the price once those cards are offered to the market again. Oracle claims the opposite of a price slump: in its first quarter, expiring GPU capacity was renewed or resold at a 20 percent premium on long-term contracts. That is Oracle's own statement, posted on X according to MarketScale, and it does not say whether the Microsoft capacity is part of it.

Availability depends on power and on the provider's schedule. The Nscale case shows a site can fail on its grid connection even when a large customer had committed. For renters, what counts is whether the contract names a fixed site, a fixed number of cards and a deadline for delivery.

On contract length it is unclear how neoclouds handle freed capacity. The reports say nothing on it, and a short term keeps you flexible. What it costs compared to a long term has to come from the provider, and we have no sourced figures.

We covered the broader situation in Neoclouds under pressure and rented GPU capacity. What is new is that the sector's biggest customer now appears as a reviewer, if only through anonymous sources.

Local inference as an alternative

Your own cards free you from renewal rounds. They replace provider risk with your own operations. According to our inventory, .180.3 holds four RTX 5090 on bare metal with 130 GB of VRAM combined, and .180.202 has another one (state of 19 August and 23 September 2026). The hub Local AI and GPU hardware collects more on cards and self-hosting. The same list records the downside: .180.3 has only 31 GB of system memory for four cards, and loading models there takes long. The files say nothing about the cost of our hosts compared with rental prices, and we do not claim a comparison here.

Whether local inference fits depends on utilization. A card that runs full all day tends to be easier to plan around when you own it. For occasional peaks, renting stays sensible, ideally with models you can move to another provider within days.

If you rent compute or are weighing your own hardware, the services page lists what we offer for own language models on in-house hardware or rented GPU capacity.

What to settle now

For a running rental contract, three questions to the provider are worth asking: which chip generation sits behind the offer, how long the contract runs, and what happens if delivery is late. It stays open which neoclouds Microsoft keeps. The quarterly reports from CoreWeave and Nebius next week will show whether the providers address it themselves.

Further reading

• CNBC, Microsoft plans to use fewer neoclouds (6 Nov 2026): cnbc.com
• The Next Web on the Bloomberg report about Oracle (6 Nov 2026): thenextweb.com
• MarketScale with Oracle's statement (6 Nov 2026): marketscale.com