The AI infrastructure buildout is already measured in trillions. In its February 2026 report The AI Journey – From Generative to Agentic, GlobalData estimates that data centres built for AI workloads will require $5 trillion of capital expenditure worldwide between 2025 and 2030, alongside $1.5 trillion for conventional workloads. This AI bill covers land and buildings as well as graphic processing units, memory, networking, cooling and power distribution. But much of that investment is being committed before the eventual scale of enterprise demand is clear.
Near-term figures show how quickly investment is rising. GlobalData puts AI capital expenditure by five major US technology companies at $244 billion in 2024 and estimated $404 billion for 2025. The infrastructure being built now must serve workloads for years, even as model prices and computing efficiency change. Investors therefore need to know which applications will generate enough sustained use to justify it.
Agentic AI has been identified as a crucial part of that answer. An AI agent can pursue a goal through several steps, drawing on company data, using software tools, calling on other agents, and checking its results. As it works through a task, the agent may use a large language model (LLM) several times. Routine automated workflows could therefore create recurring demand for computing capacity if businesses deploy them widely.
How agentic AI drives demand for computing power
The AI Journey report maps a progression from rule-based chatbots to agents that coordinate work across departments. A basic system might retrieve a support answer. A more advanced agent could combine customer relationship data, service tickets and financial reports to manage a sales pipeline. Several agents could also collaborate on orders and inventory. These examples show why more independent agents may need access to more information, use the AI model more often, and coordinate more tasks.
LLMs process those calls as tokens, which are the units into which text is divided for computation and billing. As agents plan, act and review their work, they can generate far more input and output tokens than a one-off query. Enterprise agents are expected to become a major source of demand for model services, but this outcome still depends on adoption, because cheaper tokens and more efficient chips could partly offset the extra workload.
More inference, which is the process of running a trained model to produce an answer or action, has physical consequences. Data centre operators need server capacity, fast memory and networking, reliable electricity and cooling. GlobalData’s capacity forecast illustrates the scale of the planned expansion, while its separate Data Centers report highlights electricity demand, water use and the growing interest in liquid cooling. Power availability can determine when a project actually starts serving customers.
The returns depend on enterprise AI adoption
In GlobalData’s model, a base scenario with slower enterprise uptake produces a cumulative return on investment of about 3.2% for large language model infrastructure providers by 2030. An optimistic scenario, with stronger consumer and enterprise adoption, produces about 14.6%. Both are modelled outcomes, based on assumptions about usage, revenue and a cumulative $5 trillion capital outlay. Neither is a guaranteed return.
There are financial risks along the supply chain too, such as interlinked deals in which a chipmaker invests in cloud infrastructure providers that then buy its chips. Those arrangements can help build capacity, but if customer demand falls short, weakness could travel through several companies at once.
Enterprise readiness is another constraint. GlobalData’s July 2026 report Overcoming Barriers to Enterprise AI Adoption identifies strategy, data and technology, talent and governance as four requirements for deploying AI at scale. Agents cannot deliver reliable work if the information they need sits in incompatible systems or lacks clear ownership. Spending on infrastructure will translate into lasting revenue only when businesses can move useful applications beyond pilots.
What sector leaders should track in the AI investment boom
For infrastructure suppliers and investors, announced capital expenditure is only a starting point. Leaders should track which projects secure grid connections, when capacity becomes operational, how much is contracted and how intensively customers use it. Changes in cost per inference are important as well. More efficient models could lower costs while allowing more tasks to run, making demand harder to infer from headline spending alone.
For enterprises, the test is more direct. A deployed agent should save time, improve a decision or increase revenue after model charges, integration work and oversight are counted. GlobalData’s adoption framework gives leaders a way to examine the data, skills and governance behind those results. Its research connects demand for AI applications with the chips, facilities and power systems being built to support them.
The next phase of the boom will be judged by that connection. GlobalData’s Technology Intelligence Centre brings together analysis of agentic AI, data centres and enterprise adoption to help leaders assess where demand is emerging and where expectations may be running ahead of use.
Download the in-depth Agentic AI report for the evidence behind the investment outlook.
Bibliography
GlobalData, Strategic Intelligence: The AI Journey – From Generative to Agentic, 19 February 2026, pp. 8, 11, 16–18, 28 and 30–31.
GlobalData, Strategic Intelligence: Data Centers, September 2025, executive summary and key highlights.
GlobalData, Strategic Intelligence: Overcoming Barriers to Enterprise AI Adoption, 2 July 2026, pp. 4–6, 9 and 22–24.
