Cloud lock-in used to begin with software: databases, identity systems and proprietary services accumulated until leaving required an executive sponsor and a small migration army. AI has reversed the sequence. Enterprises now choose infrastructure according to where the accelerators are available, then tolerate whatever cloud happens to own, or lease, the building around them.
Nvidia's data-center revenue reached $115.2 billion in fiscal 2025, according to Nvidia SEC filings, meaning a single chipmaker's server business generated more annual sales than many enterprise-software portfolios assembled across decades. Microsoft said it expected to spend roughly $80 billion on AI-enabled data centers during its fiscal 2025, according to Microsoft, enough capital to reproduce the market value of a respectable public software company every few weeks. Customers still encounter queues, allocation limits and awkward conversations with cloud account teams. Scarcity, rented.
The result is a new infrastructure tier, made up of CoreWeave, Lambda, Crusoe, Nebius and a lengthening cast of regional operators, that buys accelerators, assembles high-bandwidth clusters and sells compute without asking customers to adopt an adjoining catalog of databases, office software or advertising tools. TIME's profile of CoreWeave described chip scarcity as the opening that allowed the specialist provider to expand from three data centers to 14 and sell capacity even to Microsoft.
An enterprise trusting a young provider with an expensive model-training run looks reckless until the alternative is priced: waiting, while an idle machine-learning team burns payroll and produces nothing.
CoreWeave generated $1.9 billion of revenue in 2024, according to CoreWeave SEC filings, meaning a specialist GPU landlord built a large enterprise-infrastructure business before most CIOs had settled on an AI architecture. Microsoft supplied 77% of that revenue, according to the same filings, meaning 77 cents of every sales dollar came from the company whose cloud platform CoreWeave is supposedly challenging. A strange inversion, worth sitting with.
Hyperscalers are becoming customers of the challengers because the scarce resource is no longer demand or distribution. It is energized capacity with GPUs attached. Nebius signed an infrastructure agreement worth $17.4 billion with Microsoft, according to Nebius, meaning a buyer with its own globe-spanning cloud committed a nation-state-sized technology budget to an outside compute supplier. The independent clouds are not merely stealing workloads at the edge. They are acting as overflow balance sheets for the incumbents. The engineering underneath is real: dense accelerator clusters require liquid cooling, low-latency networking, fast checkpoint storage and schedulers capable of keeping thousands of costly chips occupied, and a failed fabric link can strand an otherwise healthy rack.
But availability is not a moat. CoreWeave reported $25.9 billion in revenue backlog after its first public quarter, according to CoreWeave first-quarter results, meaning much of its future capacity had buyers before the concrete, switchgear and GPUs were fully earning revenue. OpenAI separately committed as much as $11.9 billion under a CoreWeave services agreement, according to CoreWeave SEC filings, meaning one AI developer effectively underwrote a substantial slice of the supplier's expansion. Those contracts make infrastructure finance possible, but they also replace diversified cloud economics with customer concentration, hardware depreciation and refinancing risk.
Investors are apt to value these businesses as fast-growing cloud platforms, and the balance sheet argues otherwise. Their current economics look closer to contracted power generation crossed with aircraft leasing: enormous upfront spending, valuable scarce assets, concentrated customers and returns that depend on keeping equipment occupied after the initial contract rolls off.
The hyperscalers still hold data gravity, developer ecosystems and procurement relationships that independent operators cannot recreate by stacking Nvidia boxes. When accelerator supply loosens, generic rental pricing will fall. When a new chip generation arrives, yesterday's prized cluster becomes the discounted aisle surprisingly fast.
The investable question is not whether independent AI clouds can sell scarce compute. They plainly can. It is whether they can turn temporary scarcity into durable scheduling software, customer trust and lower operating costs before capital markets discover that GPUs are machines, not magic.
CoreWeave's backlog is the test of that question: when the queue disappears, the scheduling software and the cost base are what is left to sell.