Gartner now expects that over 80% of "neocloud" providers, the GPU-rental clouds built around AI demand, will cease independent operations by 2030. That forecast, plus a rough fortnight for CoreWeave and Nscale, is the backdrop to The Decoder's analysis, which argues a shakeout is coming and that Nvidia has the most to lose. This post sets out what is reported, who said it, what is a forecast versus a fact, and what it means if you rent GPUs for training or inference.
We have been tracking the pressure points one at a time: Microsoft pulling out of Nscale's Loughton site over grid power, record GPU rental prices and CoreWeave's first Vera Rubin cluster. This piece connects them.
TL;DR: the claims and who made them
| Claim | Source | Status |
|---|---|---|
| Over 80% of neoclouds may cease independent operations by 2030 | Gartner press release, Oct 14, 2026 | A forecast, conditional on current spending |
| Over $3.5 trillion of AI infrastructure spending could be exposed to impairment by end of 2030 | Gartner press release | A modeled scenario |
| Only seven to ten neoclouds survive a few years | Industry executives, per The Decoder | Opinion |
| CoreWeave lost about a quarter of market value; credit risk gauge at a record | Bloomberg, per The Decoder | Market data as reported |
| Microsoft pulled out of an Nscale Loughton deal for 23,000 Nvidia chips | The Telegraph, per The Decoder | Reported; Nscale calls the relationship strong |
| Non-hyperscaler customers were $40B, 45% of Nvidia data center revenue in the July quarter | The Decoder | Reported figure |
| Bernstein: only 35% of announced North American capacity gets built | Bloomberg, per The Decoder | Analyst estimate |
What did Gartner actually say?
Gartner's October 14 press release says that over $3.5 trillion of AI infrastructure spending could be exposed to impairment by the end of 2030. Coverage of the note reports Gartner expects roughly $1 trillion of AI infrastructure spending in 2026 and about $6 trillion more from 2027 to 2030, against no more than $3.6 trillion of AI revenue attributable to AI data centers in that period, with a 12% return on invested capital not expected until the end of 2032 at the earliest. Two caveats matter. The 80% figure is conditional on spending continuing at the current pace, so it is a scenario rather than a prediction about any named firm. And coverage notes the release does not define which companies count as neoclouds or publish the model behind its totals. Read it as a warning about economics, not a countdown.
How does a neocloud make money, and why is that fragile?
A neocloud buys or leases data center space, fills it with Nvidia accelerators, and rents the capacity back, usually under multi-year contracts to a handful of large AI customers. The model works while three things hold: customers keep signing, chips can be financed cheaply, and power arrives on schedule. Each of those is now being tested in the reporting above.
- Customer concentration. A single anchor tenant can account for a large share of contracted revenue, as the reported Microsoft share of Nscale's backlog shows. If that tenant walks, the lender's view of the whole project changes.
- Financing costs. Chips depreciate quickly while debt is long. A record credit-risk reading for CoreWeave, as Bloomberg reports, means the next round of borrowing costs more for everyone in the category, not just the weakest names.
- Power and siting. Loughton is the clearest example: a site can be leased and designed and still wait years for a grid connection.
- Differentiation. Gartner's phrase is a "lack of a differentiating business model." Renting the same GPUs as a rival leaves price as the only lever, which is how commodity markets consolidate. Software and tooling layers, which Intrator points to in his memo, are the usual escape route.
This is also why the hyperscalers are treated differently in the forecasts: they have other businesses to absorb underused capacity, while a pure-play GPU renter does not.
What happened to the stocks and the deals
The Decoder, citing The Information and Bloomberg, reports a bad stretch. Australian provider Firmus pulled its IPO. Shares of CoreWeave and smaller rivals fell about 20%, per The Information. Bloomberg reports CoreWeave lost a quarter of its market value and that a measure of its credit risk hit a record high.
On Nscale, The Telegraph reports Microsoft pulled out of a deal at Loughton, England, where it had planned to lease 23,000 Nvidia chips, with utilities not expecting power at the site until the 2030s. Nscale says its Microsoft relationship is "strong and long-term." Bloomberg reports some investors want Nscale's target market cap cut from about $30 billion to $16 to $19 billion, and The Telegraph says over 40% of Nscale's $103 billion of contracted revenue comes from Microsoft. We examined the Loughton facts in our earlier post; the broader story of Nscale's contracts also includes a 100,000-GPU Vera Rubin deal with Figure.
Server towers and a compute chip illustrating the neocloud shakeout in GPU data centers
Why would Nvidia lose the most?
Per The Decoder, customers outside the hyperscalers, neoclouds included, generated $40 billion in Nvidia's July quarter. That is 45% of data center revenue and more than double the year-earlier figure. If a wave of providers fails or cannot finance new buildouts, a large slice of that demand disappears.
The exposure is not only sales. The Decoder reports Nvidia supports neoclouds with equity investments, chip allocations, loan guarantees and customer introductions, and that a team led by Nico Caprez, Jensen Huang's son-in-law, helps them find sites with available power. It also reports Nvidia changed its loan guarantee terms after larger neoclouds rejected a first offer that would have taken a share of future revenue; the new offer has shorter terms for a smaller cut. For the financing side of this relationship, see our earlier analysis of Nvidia as the central bank of AI vendor financing.
The Decoder draws a distinction between business models. CoreWeave mostly leases its data centers, so sites could be re-leased more easily if it failed. Nebius, IREN and Nscale build their own facilities, so trouble could stall buildouts and cut Nvidia's sales.
The industry is not unified
CoreWeave CEO Michael Intrator blamed smaller rivals Firmus, Lambda and Nscale for the turmoil in an internal memo, writing that they are "trying to raise enormous amounts of capital without the execution, revenue, scale, or software solutions that CoreWeave delivers," according to The Decoder. Cirrascale CEO David Driggers said "We'll definitely see consolidation," expecting asset managers and telecom and fiber companies to buy neoclouds. These are positions from interested parties.
The oversupply argument
An unnamed neocloud executive told The Decoder that compute supply will eventually outstrip demand. The Decoder links this to Microsoft CEO Satya Nadella's late February 2025 podcast remarks expecting oversupply starting in 2027, with prices falling as capacity comes online. Bernstein's estimate that only 35% of announced North American capacity will be built points the other way: supply is as uncertain as demand. Our data center backlash map and hyperscaler nuclear deals guide show why: power, siting and local opposition are hard limits.
What does this mean for people renting GPUs?
You do not need to panic, but you should plan for counterparty risk.
- Keep workloads portable. Containerize, avoid provider-specific orchestration where you can, and test failover to a second cloud.
- Read continuity terms. Check what happens to reserved capacity, data and egress if a provider is acquired or enters restructuring. Consolidation is more likely than shutdown, so change-of-control clauses matter.
- Do not read the forecast as a price signal yet. Spot prices for top GPUs have been climbing; see the B300 rental price claim and what is verified. Oversupply, if it comes, is a 2027-and-later story in the sources above.
- Favor providers with diversified customers. Concentration, like Nscale's reported Microsoft share, is the risk to ask about.
- Own less lock-in with inference. Open-weight models you can host anywhere reduce the cost of moving.
What is not established
No neocloud has gone bankrupt in this reporting. Gartner's number is a modeled outcome. Several figures come second-hand through The Decoder from The Information, Bloomberg and The Telegraph, which we have not independently opened; verify against the originals before relying on them. Stock moves are market noise, and Nvidia's loan guarantee details come from press reports, not a filing we have reviewed.
What to watch next
- Whether Firmus or Nscale revive their IPO plans and at what valuation.
- Any further hyperscaler lease cancellations.
- Nvidia's next earnings call, for commentary on non-hyperscaler customers.
- Consolidation deals involving telecom, fiber or asset-manager buyers.
Related reading
- Microsoft pulls out of Nscale's Loughton datacentre
- Nvidia B300 rental prices: claimed vs verified
- CoreWeave deploys first Vera Rubin cluster
- Nvidia as central bank of AI vendor financing
- Figure and Nscale 100,000 GPU deal
- AI data center backlash map
- Hyperscaler nuclear deals
Figures reflect reports and the Gartner release as of October 19, 2026 and may change.