Microsoft Neocloud Pullback Piles Pressure on CoreWeave, Nebius
Microsoft has started to unpick its dependence on rented AI compute, and the neocloud sector is feeling the strain. The company is winding down a 2023 contract to lease capacity from Oracle and reviewing the wider web of deals through which it has committed more than $80 billion to specialist providers, according to reporting by Bloomberg. For shareholders in CoreWeave and Nebius Group, the Microsoft neocloud pullback turns one lost contract into a broader question: how durable is a model built on hyperscaler demand that the hyperscalers are now trying to bring in-house? Microsoft sits at the centre of it as both the sector's largest customer and its emerging rival.
What has Microsoft actually changed?
The immediate trigger is Oracle. Microsoft has declined to renew a contract, first struck in 2023, under which Oracle supplied servers to run an AI-assisted version of Bing. Bloomberg reported on 6 November 2026 that Microsoft had been switching off that capacity over recent months, partly because the rented servers ran older Nvidia chips that no longer met its specifications. The financial terms were never disclosed.
The Oracle decision sits inside a larger review. Microsoft has committed more than $80 billion to rent capacity from neocloud providers including CoreWeave, Nebius and Nscale, and it is now reassessing those arrangements. Over the same period it tightened oversight of projects run by Nebius, IREN and Nscale, declined an option to expand a Nscale site in Texas and stepped back from a Nscale development in the UK. Microsoft is not abandoning rented compute. It is consolidating the number of suppliers it works with and steering more spending towards its own data centres.
Why is Microsoft building its own chips?
The strategic logic sharpens alongside Microsoft's silicon roadmap. On 12 October 2026 the company detailed Maia 300, the third generation of its in-house AI accelerator, with production deployments scheduled for 2027. Microsoft says the inference-focused chip will deliver about 40% more tokens per megawatt than leading GPU-based systems, and that a full 96-accelerator rack will offer up to seven times the throughput of its predecessor. Scott Guthrie, the executive vice-president for cloud and AI, has positioned it as the primary engine for Microsoft's own inference workloads.
Crucially, Maia 300 runs only inside Microsoft's data centres. As production ramps through 2027, every workload the company moves onto its own cheaper silicon is a workload it no longer needs to rent. That is the backdrop to the neocloud review: cheaper internal capacity weakens the marginal case for signing fresh third-party contracts.
How exposed are CoreWeave, Nebius and Oracle?
The exposure is uneven but material. Microsoft is CoreWeave's most important customer. The company's single largest client accounted for roughly 36% of second-quarter revenue, and Microsoft made up about 67% of CoreWeave's 2025 revenue. CoreWeave reported second-quarter revenue of $2.58 billion, up 112% year on year, alongside a net loss of $626 million and an interest bill of $640 million, so any softening in Microsoft demand lands on an already stretched income statement.
Nebius is tied to Microsoft through a contract worth approximately $17.4 billion, running to 2031 and rising to as much as $19.4 billion if Microsoft takes up further capacity. Signed in September 2025 and anchored by a new data centre in Vineland, New Jersey, that single deal represents a substantial share of the roughly $3.35 billion in revenue Nebius is expected to generate this year.
Oracle is the most insulated of the three. Its cloud business booked $18 billion in revenue for the year to May 2026 across customers that include OpenAI and Meta, and the lapsed Microsoft arrangement was never quantified publicly. Tellingly, Oracle has been re-letting four-year-old GPUs at a 20% premium to their original price, a sign that demand for compute remains tight even as Microsoft steps back.
What does the pullback mean for neocloud financing?
This is where the pressure concentrates. Neoclouds are built on debt, and their ability to raise it depends on signing long-dated contracts with investment-grade customers such as Microsoft and Alphabet. As those anchor contracts come into question, the cost of borrowing has climbed sharply. The yield on some CoreWeave bonds recently rose above 16%. For comparison, Microsoft is raising $25 billion of fresh debt at rates between 4.85% and 6.35%.
The gap illustrates a widening divide. The Wall Street Journal reported in late October 2026 that some neocloud operators could not raise as much capital as they needed to keep expanding. With hyperscalers issuing their own debt in size, and a finite pool of capital to go round, investors are demanding a steep risk premium to hold neocloud paper.
Is this a one-off or an industry shift?
The evidence points to a pattern rather than a single decision. Alphabet has described its own contracts with third-party providers as a short-term "bridging strategy" while it builds out internal capacity, echoing Microsoft's direction of travel. The near-term picture still favours the neoclouds, since the shortage of GPUs remains acute and Oracle's 20% re-letting premium shows buyers will still pay up. The open question is how long that tightness lasts once hyperscaler insourcing gathers pace.
Why This Matters to FinanceX Readers
The neocloud trade was always a leveraged wager that hyperscaler demand for outsourced compute would prove durable. Microsoft's review, Alphabet's framing of third-party deals as a bridge, and the arrival of cheaper in-house silicon all point the other way: towards outsourcing as a temporary stopgap rather than a permanent fixture. For investors, that reframes the sector around three compounding risks: customer concentration, heavy leverage and the prospect of your biggest client becoming your competitor. With some bondholders already demanding 16%, equity holders need to ask what return compensates them for sitting beneath that debt.
