Gartner forecasts massive AI infrastructure write-downs
Gartner forecasts that up to 80% of neoclouds will cease independent operations by 2030, as projected AI infrastructure costs exceed projected revenue.
By
Beth Pariseau, Lead Editor, IT Infrastructure News
Published: 14 Oct 2026
Gartner predicts that AI infrastructure providers are unlikely to recoup the trillions they plan to spend, adding another variable to an already complex cost calculus for enterprise AI.
Published this week, Gartner's new forecast estimated that some $3.5 trillion in AI infrastructure investments will ultimately be written off, triggering extreme market consolidation by 2030. This is because AI providers are expected to spend up to $7 trillion on infrastructure by 2030, which would require $8.5 trillion in revenues to cover announced capital spending, financing costs and a 12% return on invested capital. But AI providers are expected to generate only $3.5 trillion in AI revenue over the same period, according to the Gartner's calculations.
Will Sommer, senior director analyst at Gartner, said the companies most heavily affected by these trends, should they come to pass, will be a "Cambrian explosion" of neocloud providers, which are providers that arose specifically to handle AI workloads.
"We [take the] position that 80% of neoclouds will cease independent operations by 2030," Sommer said. "That doesn't mean that compute assets are just going to disappear -- ceasing independent operations may include acquisition, it may include partnerships, or it may include outright failure. Not every single company that disappears is going to fail."
Still, the disappearance of a familiar vendor into another raises significant considerations for CIOs. Weighing the potential longevity of neocloud vendors, along with how they are capitalized, could also sway an ongoing 'build vs buy' dilemma enterprises face when they weigh private AI security, data gravity and potentially lower token costs vs the complexity and costs of maintaining their own AI infrastructure in-house.
"For the first time in really a generation, the math is starting to move toward 'build' and away from 'buy' or 'partner.'" Sommer said. "It doesn't mean that 'build' is preferred across the map, but it is starting to make a lot more sense."
Build, buy – where's the ROI?
At the same time, compute resources and data storage are scarce and costly -- and will remain so for at least the near future. One of the ways some organizations split the difference is to use edge devices -- even Mac Minis and laptops with GPU hardware -- for at least some AI workloads, Sommer said.
"Running relatively smaller models on local compute, whether that be a laptop that has a relatively souped-up graphics processing unit and CPU or an on-premises server or some combination of those, it's going to be use-case dependent. But certainly we see more opportunity for those types of [deployments], given the nature of AI and the inherent scarcity and cost pressure [of infrastructure]," he said.
But evaluating whether to use local hardware isn't just a matter of comparing the cost of equipment with the cost of using a cloud service. There's also the assessment of likely utilization of an on-premises hardware investment and whether lighter-weight models will adequately support a specific application, according to one IT practitioner.
Frontier models are better at generating fast responses than the kind of smaller, lighter-weight models that are practical to run on-premises. "The only way to get a good ROI on local models run on your infrastructure is saturating it around the clock," said Steve Koelpin, an AI observability engineer who runs AI models on his own hardware.
"So, this approach makes sense when you're not directly in front of the model chatting with it but rather letting it run long independent agentic workflows that don't require babysitting," Koelpin said.
Attaining real ROI for AI applications and infrastructure remains highly elusive as enterprise AI matures, according to recent market research from other analyst firms.
For example, a third-quarter survey of more than 500 enterprise and technology business leaders conducted by HyperFrame Research found that 48% of AI and ML projects reached production, but only 27% also met their original business or ROI objectives. That represents an increase compared to the first quarter's survey, which saw just 23% of projects meet their original business or ROI objectives.
We have a long way to go before organizations figure out they're spending too much money.
Rob Strechay,Lead analyst, VB Intelligence
In another third-quarter survey of 375 companies by VentureBeat Research, a similar number, 40%, of respondents said they'd tracked AI ROI rigorously. Of that number, only about 10% said they had their finance department's approval and buy-in on that ROI number in a smaller, September portion of the survey with 102 respondents.
Only 21.6% of respondents to VentureBeat's September infrastructure survey reported rigorously tracking AI infrastructure ROI. Of that group, just 45.5% had finance sign-off, which was equivalent to only 9.8% of all respondents, said Rob Strechay, lead analyst at VB Intelligence, who conducted the survey.
"That's pretty telling that nobody's at that level of maturity yet," Strechay said. "But also, we have a long way to go before organizations figure out they're spending too much money."
Market research differs on enterprise neocloud interest
Further muddying the investment outlook for CIOs is the fact that market studies sharply differ on how widely used neoclouds are among enterprise companies. HyperFrame's survey found that neoclouds were in use by just 2% of respondents, with a projected growth to 8%; by comparison, AWS was in use by 59% of respondents and Azure by 51%.
VentureBeat's study, however, found that 42.1% of respondents were interested in doing more with neoclouds.
Neoclouds, which sprang up to meet outsized demand for AI infrastructure, play an important role in cases where digital sovereignty is a priority for enterprise buyers. Even if the industry dramatically consolidates by 2030, that leaves plenty of short-term opportunities for enterprises to source AI infrastructure, Strechay said.
"It's a question CIOs should be asking of their neoclouds: how are they capitalized, what does that mean over the long term?" Strechay said. "I think most of these companies that are going to the neoclouds, it's such a two- or three-year deal kind of thing. … It's only 4 years away, but 2030 feels like it's 10 years away in AI terms."
Hyperscalers, meanwhile, are much more diversified than neocloud vendors. Still, they are not immune to potential write-downs on huge infrastructure investments, Sommer said.
"Whether or not all of the hyperscalers end up winning this market, I think it is still TBD," he said. "It may be that some of these hyperscalers or near-hyperscalers compete, don't win sufficient market share to remain a GPU-as-a-service provider, and pivot to something else."
While the big three clouds can absorb AI infrastructure losses, some second-tier cloud infrastructure vendors might also be subject to consolidation, Strechay said.
"Some of them are private-equity-backed as well, and so their problem is how much money they have to go after and take [another cloud] private," Strechay said. "As the price of these AI data centers comes down, there has to be consolidation. There's just no way all of these will last."
Beth Pariseau, senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism. Have a tip? Email her or connect on LinkedIn.
