The cloud industry has sold infrastructure in neat, stacked layers for close to two decades: storage here, compute there, platform services on top. According to a new report from research firm Omdia, however, that structure is starting to come apart, and large language models are the reason why.
The report, Global AI Cloud Stack 2026, argues that a single API call to a modern AI model now does the work that used to require four or five separate layers of cloud architecture stitched together. Instead of buying storage, compute, and platform tools as distinct products, businesses increasingly just call a model and let it handle the rest internally. As a result, the old boundaries between layers collapse into something closer to one function.
Billing has followed a similar path. Companies used to pay for each layer of infrastructure separately, but now, Omdia says, tokens have become the dominant unit enterprises pay for, which strips much of the standalone commercial value out of the old layered model. In other words, if a business is charged by the token, it stops caring how many technical layers sit underneath.
There’s a labor angle too. Tools like Salesforce’s Agentforce and Microsoft’s Copilot increasingly act less like software features and more like workers handling entire job functions on their own. Anthropic, for example, has released both a coworker-style product and a separate one built for developers to manage. As machines take on more autonomous decision-making, Omdia argues, the old assumption that humans directly control every layer of the stack starts to break down as well.
To track the shift, Omdia has redrawn the market into three categories: AI Cloud Infrastructure, Model-as-a-Service, and Agent-as-a-Service. According to the firm, the AI Cloud Infrastructure market reached $65.74 billion in 2025, with North America holding the largest regional share at roughly 46%.
Raymond Zhan, a senior analyst at Omdia, said the industry is moving toward treating models themselves as a kind of resource. Meanwhile, he added, as infrastructure vendors push further up the stack, they’ll need real return-on-investment proof in specific industries before the application layer’s value becomes fully clear. China, he noted, needs its own separate market analysis given its different billing norms and supply chains.
