Artificial intelligence has no shortage of investment. Technology companies are committing hundreds of billions to models, chips, data centres and energy infrastructure as AI adoption accelerates across the enterprise. But for investors and decision-makers, the scale of that commitment is making the question of return increasingly important: how much economic value must AI ultimately create to justify the infrastructure being built around it?

There is little doubt that artificial intelligence has moved beyond the experimental stage. Generative AI is embedded in productivity software, software development, customer service, cybersecurity, research and an expanding range of enterprise workflows. AI agents are beginning to take on more complex tasks, while the largest technology companies are building the infrastructure required to support increasingly capable models and much greater levels of inference.

The investment required to sustain that expansion is substantial. The Bank for International Settlements estimates that the five largest global technology companies will spend more than $1 trillion on AI-related investment across 2025 and 2026. Looking further ahead, total global AI investment could reach around $4 trillion by 2030 as spending expands across semiconductors, data centres, electricity generation and the wider infrastructure required to support AI workloads.

These figures do not necessarily mean that AI is overvalued or that the investment will prove excessive. They do, however, make the economic test facing the technology considerably more demanding. AI is no longer being asked simply to demonstrate that a model can improve productivity or automate a particular process. The value created across the economy will ultimately have to support an infrastructure build-out measured in trillions, shifting the ROI discussion from the cost of individual AI tools towards the economics of the infrastructure supporting them.

AI is becoming an infrastructure investment

The first phase of generative AI was largely experienced through software. ChatGPT, copilots and enterprise assistants made the technology appear almost weightless to the end user. A prompt was entered and an answer appeared, while the substantial computing infrastructure behind that interaction remained largely invisible.

That infrastructure is becoming increasingly difficult to ignore. Google announced this month that it intends to invest around €13 billion in AI and cloud infrastructure in Finland, including three new data centres. The programme also includes a long-term agreement to purchase electricity from a planned nuclear power project, illustrating how closely the economics of AI are becoming connected to energy infrastructure.

Microsoft is pursuing similarly ambitious expansion, with reported plans to increase its global data-centre capacity to around 38 gigawatts by 2032, more than three times its current footprint. These investments extend far beyond the cost of training the next frontier model. They support a computing environment that includes accelerators, networking, storage, cooling, power generation, grid connections, land and increasingly specialised data-centre infrastructure.

Inference creates another persistent source of demand because deployed models continue consuming computing resources every time they are used. As usage grows and autonomous agents undertake longer and more complex tasks, the relationship between adoption and infrastructure consumption becomes increasingly important. The economics of AI consequently begin to resemble those of an infrastructure industry as much as those of conventional software.

Productivity is only one side of the calculation

The case for this investment rests on the expectation that AI will generate economic value on a comparable scale. There is already evidence that it can improve productivity in specific tasks, accelerate software development, reduce the time required for research and analysis, and automate parts of established business processes. The difficulty comes when organisations attempt to translate those improvements into enterprise-wide financial returns.

Saving an employee an hour does not automatically create an hour of additional revenue, while producing content more quickly does not necessarily increase sales. Automating a process may reduce costs, but only if the organisation changes the surrounding workflow sufficiently to capture the saving. AI adoption can therefore rise rapidly without every deployment producing an equally measurable return.

This distinction matters because enterprise experimentation has been relatively inexpensive compared with the infrastructure now being constructed to support it. Buying licences for an AI assistant is one investment decision; financing the computing capacity required for hundreds of millions of people and increasingly autonomous systems to use AI continuously is another. As the capital intensity of the technology increases, adoption alone becomes a less satisfactory measure of economic success.

The full cost of AI is getting larger

Computing infrastructure is also only part of the investment required to deploy AI effectively. Enterprises increasingly need governance, security, identity and access controls, monitoring, model evaluation, data management and human oversight. Regulatory requirements add another layer, particularly in Europe as the AI Act moves from legislation towards implementation and enforcement.

Cybersecurity is becoming especially important as frontier models acquire capabilities that can be used defensively while also potentially accelerating vulnerability discovery, exploit development and other offensive activities. Organisations deploying autonomous agents must consider not only whether those systems improve productivity, but what they are authorised to access, which actions they can perform and how those actions are monitored.

These requirements should not be interpreted simply as additional friction surrounding AI adoption. They are part of the cost of operating increasingly capable systems responsibly at scale. A realistic enterprise ROI calculation therefore has to extend beyond comparing the price of an AI service with the number of employee hours it appears to save. Integration, infrastructure, security, governance, training and organisational change all sit on the cost side of the equation alongside the productivity improvements, cost reductions or new revenues the technology may produce.

As AI moves deeper into core business processes, these surrounding costs are likely to become both more visible and more material to investment decisions.

Investment is running ahead of proven return

Large infrastructure investments made ahead of proven demand are not unique to artificial intelligence. The internet provides an obvious historical comparison. Enormous amounts of capital were invested in telecommunications infrastructure during the late 1990s as companies anticipated explosive growth in internet traffic. Some of the businesses financing that expansion failed when expectations ran ahead of revenues, but much of the infrastructure they created subsequently became essential to the digital economy.

AI could follow a very different path, but the comparison illustrates an important distinction between being right about a technology and being right about the economics of investing in it at a particular moment. The Bank for International Settlements has begun drawing attention to this issue, not because it believes AI lacks economic potential, but because the scale and concentration of investment could create financial vulnerabilities if expected returns take longer to emerge or prove smaller than markets currently anticipate.

This is also why reducing the discussion to whether AI is experiencing a bubble is of limited value. A technology can fundamentally transform an economy while individual investments made during that transformation still fail to generate acceptable returns. The commercial success of artificial intelligence and the success of every company financing AI infrastructure are not the same proposition.

Quantum and AI face different versions of the same problem

The contrast with quantum computing is particularly revealing. Quantum remains relatively early in its commercial development, requiring significant capital to build fault-tolerant systems, manufacturing capability and supporting infrastructure before a large commercial market has been demonstrated. Its challenge is whether investment can be sustained long enough for the technology to produce meaningful economic value.

AI approaches the problem from the other direction. Adoption has happened remarkably quickly, commercial services already exist and organisations across almost every sector are experimenting with the technology. Yet the infrastructure required to satisfy those expectations is expanding so rapidly that the level of economic value needed to justify the investment continues to increase.

In that sense, quantum needs sufficient capital to reach a mature market, while AI already has a substantial market but must demonstrate that its value can justify the capital flowing into it. The comparison matters because technological adoption can easily be confused with economic return. Millions of users, increasing inference volumes and widespread enterprise deployment demonstrate demand, but they do not by themselves establish that the aggregate return will exceed the cost of providing the infrastructure.

The next AI benchmark may be economic

For the past several years, competition in artificial intelligence has been measured primarily through capability. Models have been compared through benchmarks, context windows, reasoning performance, multimodal capabilities and increasingly sophisticated agentic behaviour. Infrastructure has subsequently become another competitive measure, with access to accelerators, data-centre capacity, energy and capital influencing how quickly providers can train models and how widely they can deploy them.

Economic performance may increasingly need to sit alongside those technical measures. For technology providers, that means demonstrating that AI services can produce sustainable revenues and margins rather than simply rapid adoption. For enterprise decision-makers, it means identifying where AI produces genuine productivity, cost reduction or new revenue rather than deploying it primarily because competitors are doing the same. For investors, it requires distinguishing between confidence in AI as a transformative technology and confidence that individual infrastructure investments made during its expansion will generate an acceptable return.

Artificial intelligence may prove to be one of the most economically important technologies of the coming decades, and questioning its return on investment does not contradict that possibility. The scale of investment now being made makes the question unavoidable. AI has already demonstrated widespread demand and adoption; the next stage will be demonstrating that the economic value created is commensurate with the extraordinary amount of capital being spent to deliver it.