By rights, hyperscalers like AWS, Azure, and Google Cloud should be the absolute winners in the AI compute business. They have the most data centers, the strongest engineering capabilities, and the deepest customer relationships. Yet over the past two years, it has been CoreWeave and Nebius, starting from scratch, that have surged to valuations in the tens of billions of dollars. CoreWeave’s 2026 revenue guidance is $12.4 billion to $13.2 billion, with an order backlog of $104 billion; Nebius grew 454% year over year.
Behind this contrast lies a very simple truth: hyperscalers don’t have enough compute for themselves.
Hyperscalers now play two roles at once. On the one hand, they are “compute wholesalers,” renting GPU clusters to external customers; on the other hand, they are also “AI super-consumers,” using compute to train models for their own products.
Microsoft has to build Copilot and also support OpenAI’s training; Google has to train Gemini; Amazon has to train Nova and also provide compute to Anthropic. These internal demands are bottomless, and they always take top priority. External customers want to rent GPUs? Sorry, get in line.
So a strange market gap emerged—the customers with the strongest demand are precisely the ones that cannot wait.
OpenAI, Anthropic, xAI, and other frontier labs sign contracts worth tens of billions of dollars in compute, demanding delivery within months, the latest-generation GPUs, and the entire cluster tied together as one large block. Hyperscalers’ schedules cannot accommodate that, so they have to look elsewhere. That is the first reason Neo-clouds came into being: a market gap.
The second advantage of Neo-clouds is that they are “pure.”
They do not do databases, SaaS, or AI applications; they focus on exactly one thing—buying GPUs, building data centers, assembling clusters, and renting them out. This extreme focus allows them to deploy far faster than hyperscalers.
A hyperscaler data center project can easily take two years from start to finish, while a Neo-cloud can do it in a dozen months or even faster. For customers racing to keep up with AI, that is enormously valuable.
Third, Neo-clouds’ capital structure allows them to be more aggressive.
CoreWeave turns GPU clusters into collateral and raises billions of dollars in debt financing; Nebius secured long-term contracts from Microsoft-tier customers and then used those contracts as financing support. They are not pursuing profitability; they are pursuing rapid expansion. Hyperscalers cannot play this way, because they have to maintain overall profit margins.
What is interesting is that hyperscalers themselves are placing orders with Neo-clouds. Microsoft has signed large contracts with CoreWeave, and Amazon also buys compute from multiple Neo-clouds. Why? Because even they cannot absorb all demand internally.
The entire AI compute market is a hybrid structure of “hyperscaler self-build + Neo-cloud fill-in,” rather than a case of hyperscalers swallowing everything on their own.
From an investment perspective, Neo-clouds are an extremely leveraged segment. They are highly sensitive to GPU prices, utilization rates, and long-term customer contracts. If GPU prices fall, if customers default, or if demand for the next generation of GPUs weakens, Neo-cloud cash flow will tighten immediately. That is why their stock prices have much greater elasticity than Nvidia’s—during a bull market they can rise twice as much as Nvidia, and in a bear market they can also fall twice as much.
To understand Neo-cloud business models is to understand where the scarcity in the AI cloud layer comes from, how long it can last, and under what conditions it will disappear.