The global cloud infrastructure services market continues to grow at a pace that just two years ago seemed hard to sustain. Global spending reached $102.6 billion in the third quarter of 2025, representing a year-over-year increase of 25%. Moreover, this marks the fifth consecutive quarter with growth above 20%, a sign that demand is not only resisting but also reconfiguring around a phenomenon that now influences budgets, planning, and strategy: Artificial Intelligence moving from testing to large-scale deployments.
This data offers a broader interpretation beyond volume: growth remains “stable” as companies leave the experimentation phase behind. In practice, the cloud is no longer measured solely by raw power or model catalogs, but by its ability to support enterprise applications with multiple models, reliable operations, and agents functioning in real environments, ensuring business continuity and regulatory compliance.
Hyperscalers account for 66% of spend… but the focus is shifting
In that third quarter of 2025, AWS, Microsoft Azure, and Google Cloud maintained their top positions in the ranking and collectively captured 66% of global cloud infrastructure spending. Among the three, they registered a 29% year-over-year growth, reinforcing a well-known reality: the core of the market remains in the hands of the major hyperscalers.
The novelty lies in the type of competition. The research highlights a shift in priorities: less obsession with incremental improvements in model performance “by themselves” and more pressure to offer platform capabilities that enable robust deployment and operation of Artificial Intelligence, especially for applications that combine multiple models, tools, and coordinated agents.
Here, a concept is gaining traction in executive committees: multi-model as a production requirement. The idea is straightforward: if business depends on generative workloads, organizations want the ability to switch models based on cost, latency, risk control, or availability. In other words, resilience is no longer just infrastructure; it now includes model strategy.
AWS accelerates up to 20% and reports its best quarter since 2022
AWS led the market with a 32% share and a 20% year-over-year growth, its best performance since 2022. Two factors explain part of this acceleration: improved supply restrictions for compute resources and additional demand associated with its relationship with Anthropic.
An indicator of how intense the moment is can be seen in AWS’s reported backlog at the end of the quarter: $200 billion. This figure suggests future demand visibility and ongoing pressure on capacity.
On the platform side, AWS is advancing with Amazon Bedrock, which has expanded both its model choice and tools for enterprise environments. The report mentions support for Claude 4.5, 18 managed open-weight models, and improvements in capabilities like Guardrails and data automation. The message aligns with the trend: it’s not just about “having models,” but about putting safeguards, control, and operational mechanisms in place.
Meanwhile, AWS used re:Invent 2025 to bolster its end-to-end enterprise stack, from models to agents and automation, with the announcement of the Nova 2 family, Nova Act, and Nova Forge, among others. Additionally, it continued expanding regional presence with the launch of the Asia Pacific (New Zealand) region in September, with three availability zones, tied to data sovereignty and latency-sensitive workloads.
Azure grows 40% and emphasizes Foundry and agent orchestration
Microsoft Azure maintained second place with a 22% share and a 40% year-over-year growth, reflecting strong enterprise demand for Artificial Intelligence. In October, Microsoft renewed its partnership with OpenAI, strengthening the development and deployment relationship on Azure.
The platform component most highlighted is Azure AI Foundry, which continues expanding its model ecosystem. The report notes support for frontier models like Claude Opus 4.5, Claude Sonnet 4.5, and Haiku 4.5, serving over 80,000 customers with access to more than 11,000 models. This message is targeted at a very specific buyer: those who have already understood that their strategy cannot depend on a single model provider.
The decisive step in “production” comes with agents. In October, Microsoft introduced Microsoft Agent Framework, designed to build and orchestrate multi-agent systems. The report cites use cases like KPMG, which is applying it to improve audit processes—an example of how large consultancies aim to turn Artificial Intelligence into an operational advantage, not just a demo tool.
Azure’s expansion also follows the infrastructure map: Microsoft announced plans in November to extend its cloud region in Malaysia and launch a new data center region in India in 2026. As the market shifts toward production workloads, geography regains importance: data sovereignty, latency, and local capacity are key considerations.
Google Cloud rises to 11% with increasing backlog and enterprise AI traction
Google Cloud maintained its position as the third player globally, with an 11% share and a 36% annual growth. According to the report, this progress is primarily driven by its enterprise Artificial Intelligence offerings, with quarterly revenues in this segment totaling “several billion dollars.”
The backlog also grew significantly: as of September 30, Google Cloud reported $157.7 billion, up from $108.2 billion in the previous quarter. In infrastructure markets, backlog is often a more useful compass than weekly noise: it indicates signed contracts and committed demand.
In terms of products, Google continues expanding Vertex AI Model Garden, adding new models including multimodal variants of Gemini 2.5, as well as Kimi K2 Thinking and DeepSeek-V3.2. In October 2025, it launched Gemini Enterprise, a platform combining the Gemini family with enterprise agents, no-code tools, and security and governance packages—crucial when deploying Artificial Intelligence outside the laboratory.
From “benchmark” to “runbook”: operating agents in the real world is more challenging than expected
The research points to a reality many teams have already uncovered firsthand: deploying agents in production is not just about more capable models. There’s a lack of standardization and “building blocks” to ensure continuity, customer experience, and compliance simultaneously. That’s why hyperscalers are increasing investments in “build and run” capabilities—tools to build, deploy, and operate agents in a repeatable manner.
This trend is reflected in launches like AWS AgentCore and Microsoft’s Agent Framework, both aiming to make agent management governable: policies, observability, cost control, and consistent operations. The stage has shifted: testing is no longer enough; now the focus is on operating.
What is included in “cloud infrastructure” in these figures
The report adopts a broad definition of cloud infrastructure services, including bare metal as a service, IaaS, PaaS, container as a service, and serverless, provided they are hosted by third parties and consumed via the Internet. This nuance matters because it groups components that, from the client’s perspective, serve the same goal: run critical workloads with elasticity, security, and a clear operational model.
In summary, the market is expanding because demand is no longer aspirational—it’s operational. Artificial Intelligence is beginning to behave like another enterprise workload… only much more demanding in compute, data, and governance.
Frequently Asked Questions
Why is multi-model support becoming a requirement for deploying Artificial Intelligence in production?
Because companies seek resilience, cost control, and flexibility: being able to switch models based on price, performance, compliance, or availability reduces operational risk and vendor dependency.
What does it mean that the cloud market moves from “experimentation” to “production” in 2025?
It indicates that investments are shifting away from pilots toward stable deployments: governance, observability, business continuity, security, and the operation of scalable agents and generative applications.
What is the hyperscaler backlog, and why is it used as a demand indicator?
It’s the committed order portfolio (signed contracts or secured future revenue). When it increases, it often signals growth visibility and sustained demand, even if the quarter has some short-term fluctuations.
What services are included when measuring “cloud infrastructure” in these reports?
Typically, it encompasses bare metal as a service, IaaS, PaaS, managed container services (CaaS), and third-party hosted serverless offerings.
via: omdia