A new report from analyst firm Omdia, titled Global AI Cloud Stack 2026: Rethinking Cloud in the Agentic AI Era, concludes that the rise of large language models since 2023 has fundamentally disrupted the layered architecture that cloud-native computing has relied on for years. The findings have direct implications for hosting providers, infrastructure vendors, and any organization planning its AI cloud strategy.
Three Forces Breaking Down Vertical Architecture
Omdia identifies three specific pressures dismantling the traditional cloud stack:
- Cross-layer model calls. A single model API call can now perform functions that previously spanned four or five distinct cloud-native layers. Those layers are being absorbed into functional components within the model itself rather than remaining independently maintained tiers.
- Token-based billing. Enterprise procurement is shifting from paying for each discrete layer to paying per token. This change strips the original layered structure of much of its independent commercial value, reducing layers to supporting technical components rather than separately deliverable products.
- Machines as labor. Platforms such as Salesforce Agentforce, Microsoft Copilot, and Anthropic’s agent offerings are positioning AI as a workforce participant rather than a tool. As human control over automated processes loosens, the control-plane logic underpinning cloud-native design weakens alongside it.
Two Camps of AI Adoption
Omdia also distinguishes two broad approaches enterprises are taking toward AI model consumption. The first, which the report calls the “API-first” camp, prioritizes speed and cost efficiency by using off-the-shelf model capabilities without deep customization. These organizations typically deploy models for tasks such as customer service chatbots or basic content generation.
The second group, described as “Agent-centric,” focuses on orchestrating complex, domain-specific workflows. A supply chain agent, for example, might call a model API to analyze market trends, trigger inventory adjustments, coordinate with suppliers, and produce stakeholder reports, all autonomously. In this model, the AI model is one cognitive component among several rather than the entire solution.
A Revised Three-Layer AI Cloud Framework
In response to these shifts, Omdia has redefined the AI cloud technical architecture into three distinct layers:
- AI Cloud Infrastructure (Infra): Covering multi-tenant AI IaaS and physically dedicated bare-metal AI services (BMaaS).
- Model-as-a-Service (MaaS): Split into Model Production Services and Model Consumption Services.
- Agent-as-a-Service (AaaS): Encompassing both Agent Platform Services and Agent Labor Services.
Market Size Figures for 2025
Omdia put the global AI Cloud Infrastructure market at $65.74 billion in 2025. Within that figure, AI IaaS accounted for $24.79 billion, led by North America at 45.7%, followed by China at 22.3% and Europe at 15.4%. The AI BMaaS segment was larger at $40.95 billion, with North America representing 54.5% of that total.
The Model Production Service segment reached $18.54 billion, with model development and training making up 63.5% of the category. North America led regionally at 43%, followed by Europe at 16.8% and China at 14.8%.
What This Means for Hosting and Infrastructure Professionals
Omdia’s senior principal analyst for Cloud and AI notes that as the resource layer shifts upward through the stack, traditional infrastructure vendors gain the ability to compete at the application layer for the first time. For hosting providers, this suggests that the competitive boundary between infrastructure and software is continuing to erode, and that positioning around token delivery and agent orchestration may matter as much as raw compute capacity in the near term.