What did Microsoft change about its neocloud deals?
Microsoft is narrowing the list of outside AI cloud suppliers it treats as partners, according to people familiar with the matter cited by CNBC on 6 November. Those people said it is still evaluating which providers to keep and has not decided which contracts to end.
The change is not one cancellation. It is a shift in how Microsoft buys compute: more capacity in its own data centres, fewer external suppliers, and closer scrutiny of delivery dates.
Microsoft has signed deals in recent years with CoreWeave, Nebius, Chronoscale, IREN, Lambda and Nscale, as well as with Oracle. A spokesperson said the strategy is "built on flexibility and choice" across owned data centres, leased facilities and third-party partnerships.
Two people told CNBC that Microsoft opted not to renew a deal with Oracle. Bloomberg reported the Oracle wind-down first, and Reuters later summarised it. Microsoft also declined an option to more than double its Texas agreement with Nscale, Reuters reported, and walked away from anchoring Nscale's Loughton project in England.
That option decision was made several months ago, before Nscale filed for its IPO in September, and the existing Texas contract was not affected, according to Reuters.
IREN separately moved more than $4 billion of annualised revenue capacity out of 2026 and into early 2027. The trigger was a Microsoft request to reconfigure and upgrade network equipment at a site built to serve it.
Three decisions, one direction. Microsoft is trading breadth for control at the point where AI capacity is delivered, financed and paid for.
What is a neocloud, and why did Microsoft depend on them?
A neocloud is an AI-focused cloud provider that rents out graphics processing power, usually under multi-year contracts. It typically offers a narrower set of services than Amazon Web Services, Microsoft Azure or Google Cloud, and it often funds buildings and chips with debt.
Microsoft's turn to these companies began after ChatGPT launched in late 2022. Demand for Nvidia GPUs outran the capacity inside Microsoft's own data centres, and renting was the fastest way to add compute without waiting years for new buildings and power connections.
CoreWeave was a private startup with roots in cryptocurrency mining when Microsoft signed up. When CoreWeave went public last year, it disclosed that 62% of its 2024 revenue came from Microsoft. That single number explains both the neocloud model and its fragility.
Renting gave Microsoft speed and flexibility. It also made several young companies dependent on one large customer, and it gave Microsoft a growing set of contracts to manage as chips aged and its own footprint expanded.
The market has changed around those deals. OpenAI, once an Azure-only customer, now buys cloud capacity from Amazon, Cerebras, CoreWeave, Google and Oracle, according to CNBC. Microsoft is no longer the only route to frontier-model compute, and its incentive to hold every rental agreement has weakened.
Scale is another factor. A recent industry review counted 323 neoclouds, CNBC reported, and experts have predicted that many will not survive independently. When supply was scarce, breadth helped. When capital gets expensive, breadth becomes a coordination cost.
Why is Microsoft reducing its supplier list now?
The financial backdrop matters. Microsoft spent $45.2 billion on capital expenditure and finance leases in the quarter ended 30 September 2026, up about 30% from a year earlier but below the $50.67 billion analysts expected, according to CNBC. Azure revenue grew 48% to $33.17 billion.
On the earnings call, finance chief Amy Hood said some third-party capacity deliveries expected in the first quarter had slipped. She tied the delays to GPU-as-a-service providers and said the capacity had shifted into the second half of the fiscal year. Some GPU cloud contracts slipped into the first half of 2027.
Those delays cost Microsoft growth it had already sold. Hood nevertheless said demand still exceeds supply. That combination, strong demand and late deliveries, is exactly the condition that makes a customer prune its supplier list.
The mix of spending has moved toward assets that pay back quickly. About 75% of capital expenditure is now tied to short-lived assets such as Intel chips and Nvidia GPUs, up from roughly 66% in the prior three quarters and about half a year earlier, CNBC reported. Free cash flow fell 27% to $18.63 billion.
Microsoft has guided to more than $50 billion of capital spending in the current quarter and about $170 billion for calendar 2027. At that level, every rented contract competes with capacity Microsoft can own and schedule itself.
Older rented capacity also loses value as models change. The Oracle hardware used an older generation of Nvidia chips, Reuters reported, and that capacity became less useful over time. A contract signed during a shortage can look expensive once newer chips arrive.
Capital markets have tightened at the same time. Neocloud shares fell by roughly half over the past month, and IREN co-CEO Daniel Roberts said the questions have moved from contract size to "whether it can be delivered, financed, and run at a profit." Microsoft is responding to the same questions, from the customer side.
What does the Oracle wind-down show?
Oracle announced the multi-year deal with Microsoft in 2023. Microsoft was to use Oracle Cloud Infrastructure to power the AI-assisted version of Bing. Financial terms were not disclosed, which makes it impossible to size the revenue at stake for either company.
Microsoft declined to renew the arrangement and has been turning off rented capacity over recent months, according to the Bloomberg report summarised by Reuters. The report said the capacity no longer met Microsoft's evolving needs and relied on older Nvidia hardware.
Oracle pushed back on the implication that demand has weakened. A spokesperson told Reuters that demand for OCI's AI infrastructure "remains strong, including for earlier-generation GPUs." That is a company position, not an audited number, but it is a useful check on the idea that the market has suddenly emptied.
Shares of Oracle and Microsoft were flat in afternoon trading after the report. Both had already come under pressure following the Democratic sweep in the US midterm elections, which raised concerns about tighter scrutiny of the AI buildout.
The episode shows how quickly a large cloud partnership can become optional. The deal was a stopgap that helped Microsoft meet demand when its own capacity was short. As newer capacity arrives, the oldest rental is the easiest line to cut.
It also shows the limit of reading contracts as permanent alliances. Many of these agreements are capacity commitments with renewal dates, not equity stakes. A decision not to renew can be made quietly, over months, without a single announcement.
How did IREN's $4 billion delay expose delivery risk?
IREN gave the clearest example of how delivery risk travels through the sector. The company said more than $4 billion in annualised revenue that it had expected in 2026 would arrive in January 2027 at the earliest, with a possible slip to February. It cited a "customer-requested change and supplier timing".
On the earnings call, Roberts identified Microsoft as the customer. He said Microsoft asked IREN to reconfigure and upgrade network infrastructure so that 200 megawatts of capacity could serve the latest frontier models. IREN had handed over data centre capacity in August, but the upgrade moved the online date.
The commercial relationship survived. The schedule did not. That distinction matters because IREN had already booked the revenue, signed the contract and told investors when the capacity would pay.
The rest of the results were strong. Revenue rose 66% from the prior quarter to $228 million, beating the $221 million analyst estimate compiled by LSEG. AI cloud revenue more than doubled to $200 million, and the net loss narrowed to $263 million from $684 million. Revenue backlog rose 22% to $20.3 billion, with about two-thirds tied to investment-grade counterparties.
IREN said the $4 billion target is fully contracted and funded. It still plans about $15 billion of capital expenditure for the rest of the fiscal year, including $9 billion in cash and committed financing and $6 billion of target financing that depends on market conditions.
The stock showed how jumpy investors have become. IREN fell 26% after the midterm elections, then rose 16% on Friday after the results. Before that rally it was down 55% over a month. A single customer request had moved billions of dollars of revenue timing, and the share price moved with it.
Roberts argued that the public market and the customer market were telling different stories. "There is a disconnect between what we're seeing in public market and what we're seeing at the customer level," he said. Both things can be true: demand is real, and the timing risk is real.
Who is most exposed, and what are the limits of the reporting?
CoreWeave is the most visible case. Microsoft accounted for 62% of its 2024 revenue, and the company reports quarterly results on Tuesday. Its shares have fallen 60% over the past month. Nebius, which reports next Thursday, is down 48% over the same stretch, according to CNBC.
IREN, Nscale, Lambda and Chronoscale sit in the same customer-concentration bracket. Oracle is exposed in a different way: losing a marquee tenant is a reputational test even if the revenue is replaceable.
The lenders, landlords and power suppliers behind these projects also carry risk. Neoclouds often sign long leases and borrow against customer contracts. If a contract is not renewed, the building does not disappear, but the revenue backing the loan can.
Customers of the neoclouds are exposed too. Startups and AI labs that reserved capacity may find delivery dates moving, even when their contracts remain valid. The IREN case shows that a supplier can absorb a customer's redesign request by pushing the schedule, not by cancelling the deal.
The reporting has limits. CNBC's account relies on people who asked not to be named. Those people said Microsoft is still evaluating which suppliers to keep. Microsoft did not identify which contracts are at risk. Oracle said demand for its AI infrastructure remains strong, and IREN said its demand outlook is unchanged.
Microsoft's own framing is a position, not a verdict. A spokesperson called the review "a normal and responsible way to manage large, complex infrastructure projects" and said the pace of new agreements "reflects what we need today." That may be accurate, but it is also what a company would say while renegotiating from strength.
The evidence points mainly to delivery timing and the cost of capital, not to a collapse in AI usage. Microsoft says demand exceeds supply. IREN says contracts are funded. Oracle says demand is strong. What has changed is who bears the risk when schedules slip.
What should customers and investors watch next?
CoreWeave reports on Tuesday and Nebius on Thursday. The useful details will be customer concentration, backlog duration, prepayments, capex funding and delivery schedules. A growing backlog means little if the capacity arrives late or the customer can walk away at renewal.
Microsoft's next quarterly update will show whether third-party deliveries land in the second half of the fiscal year as Hood guided. Watch capital expenditure against the $170 billion calendar 2027 plan, and watch for any new neocloud agreements or quiet non-renewals disclosed in filings.
IREN's January or February go-live is a concrete test. If the network upgrade slips again, the $4 billion revenue bridge moves again, and so does the financing plan that depends on it.
Nscale has its own tests. It filed to go public in mid-September and has not updated its prospectus since, CNBC reported. Investors will want to see whether Microsoft remains a significant customer, how the Texas contract develops, and who anchors the Loughton project after Microsoft stepped away.
Policy is the wild card in Texas. Governor-elect Gina Hinojosa has pledged to halt new data centre projects until the legislature passes rules for the industry, Reuters reported. A pause would not cancel existing contracts, but it would slow the next wave of capacity that neoclouds are counting on.
For buyers of AI compute, the practical lesson is to ask about delivery dates, financing and customer concentration, not just contract size. A large contract with a thinly financed supplier is a schedule risk. A smaller contract with a diversified supplier may be worth more.
The neocloud boom is not over. The phase in which hyperscalers signed any available capacity at almost any price is. Microsoft's reset is the clearest sign yet that the buyers now have the leverage, and that they intend to use it.