• Neocloud providers built billion-dollar businesses renting AI compute that the hyperscalers couldn’t supply fast enough.
• Meta and SoftBank entered the market within 24 hours, and the scarce asset is now power, not GPUs.

The neocloud sector had a strange week: its two biggest headlines came from companies that aren’t neoclouds. Earlier this week, Bloomberg reported that Meta is developing plans to sell AI computing power to outside customers under an initiative called Meta Compute. The next day, SoftBank announced SB Neo, a US subsidiary that will sell AI compute to American enterprises, including the hyperscalers themselves, from fiscal 2027. 

The market read the Meta report as a threat. CoreWeave and Nebius, the two largest listed neocloud providers and both major Meta suppliers, dropped roughly 12%, while smaller rival IREN fell about 6.5%. The reaction made sense: when your biggest customer starts selling what you sell, the growth story needs rewriting.

What a neocloud actually is

A neocloud is a cloud provider built almost entirely around one product: renting GPU capacity for AI training and inference. Where AWS, Azure and Google Cloud offer hundreds of services spanning storage, databases and enterprise software, neoclouds do one thing at speed. 

ABI Research’s roster of leading players includes CoreWeave, Nebius, Lambda, Crusoe, Nscale, Vultr and Civo, alongside converts like IREN that pivoted from bitcoin mining. The category exists because AI demands outrun hyperscaler capacity. When model developers needed clusters of tens of thousands of GPUs in specific configurations on short timelines, whoever could deliver first won the contract, and it frequently wasn’t the incumbents. 

The dependency runs deeper than most customers realise: Microsoft has routed Azure API requests to CoreWeave data centres when short on GPU capacity of its own.

Why did the money follow

The growth numbers explain the hype better than any thesis. Nebius grew first-quarter revenue 684% year on year to US$399 million and swung to a US$129.5 million adjusted EBITDA profit. CoreWeave raised its projected 2026 capex to as much as US$35 billion and disclosed a US$21 billion agreement with Meta running through December 2032, built on top of a US$14.2 billion contract signed in September 2025. 

Nebius holds its own Meta deal worth up to US$27 billion over five years. CoreWeave says nine of the ten leading AI labs now run workloads on its platform, and Lambda is preparing an IPO after landing a Microsoft deal and US$1.5 billion in funding. CoreWeave chief executive Michael Intrator’s framing of the spending was blunt: “What I’m doing is I’m building a company.”

There’s an accounting mechanism underneath the deal flow that rarely gets explained. When a hyperscaler builds a data centre, the spending lands on its balance sheet as capex. When it rents the same capacity from a neocloud, the cost is recognised as an operating expense spread over the life of the contract. 

Meta is guiding for US$125 to US$145 billion in capital expenditure in 2026 and could easily run free cash flow negative, so pushing tens of billions of additional capacity into multi-year rental agreements keeps the buildout moving without deepening the hole. Renting was never only about speed.

The fragile parts

The model’s weaknesses were visible before this week. Customer concentration is the obvious one, with Meta alone accounting for roughly US$48 billion in commitments across CoreWeave and Nebius. The financing is circular in places, since Nvidia has invested billions in Nebius, which spends its raised capital largely on Nvidia hardware. 

The debt is novel too: CoreWeave priced a US$8.5 billion facility in March, the first investment-grade GPU-backed financing, collateralised against contracted cash flows from exactly the kind of customer that just signalled it may build instead of rent. McKinsey warned last year that the neocloud model is inherently commoditised, with limited differentiation in renting out access to the same hardware everyone else buys.

Custom silicon erodes the premise from another direction. Broadcom and Google are lining up roughly 3.5 gigawatts of TPU-based capacity for Anthropic from 2027, AWS keeps scaling its Trainium chips, and every such programme chips away at the assumption that Nvidia GPU access is a moat worth paying a premium for.

The customers become the competition

Which brings us back to the week’s inversion. Meta is debating whether to offer hosted access to its AI models in the style of AWS Bedrock or to sell raw computing capacity in the style of CoreWeave, and D.A. Davidson’s Gil Luria noted the threat lands hardest on the neoclouds because they rely on Meta for growth and, in his words, “Meta may not need them anymore.” 

SpaceX has been running the same playbook, leasing capacity at its Colossus facility to Anthropic for a reported US$1.25 billion a month and to Google for US$920 million a month. SoftBank’s Junichi Miyakawa called SB Neo his company’s “second founding,” and people familiar with the plans told Japan Times the US business could triple or quadruple the telecom unit’s annual operating income.

What SB Neo makes explicit is where the contest has moved. SoftBank’s pitch rests on a claimed edge in securing power, anchored by a planned campus on US Department of Energy land in Ohio scaling toward 10 gigawatts, with the first 800MW phase due in 2028. CoreWeave and Nebius are each sitting on 3.5 gigawatt contracted power pipelines and racing to convert them to active capacity.

IDC’s Dave McCarthy calls it a bifurcating market, and the split he describes is between those who own energised, grid-connected capacity and those who merely own chips. GPUs turned out to be the easy part. Anyone with capital can buy them, which is precisely the problem.

Whether the arrival of Meta, SoftBank and SpaceX as sellers marks the market maturing or the margins compressing depends on a question nobody can answer yet: does contracted capacity come online faster than AI compute demand grows? Enterprises buying that compute should hope the sellers keep multiplying. The sellers, presumably, hope otherwise.

Want to experience the full spectrum of enterprise technology innovation? Join TechEx in Amsterdam, California, and London. Covering AI, Big Data, Cyber Security, IoT, Digital Transformation, Intelligent Automation, Edge Computing, and Data Centres, TechEx brings together global leaders to share real-world use cases and in-depth insights. Click here for more information.

TechHQ is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.