The AI Economy: Who Actually Makes Money From AI?

The first layer: compute

At the centre of the AI investment cycle is compute. Training and running increasingly capable models requires vast quantities of specialised processing capacity, which has created extraordinary demand for advanced accelerators and the surrounding infrastructure needed to operate them.

The obvious beneficiaries are the companies supplying the processors themselves, but focusing only on the headline chip designer misses a significant part of the economics. Modern AI data centres require high-bandwidth memory, advanced packaging, networking equipment, power-management systems and increasingly sophisticated cooling infrastructure. A processor cannot generate revenue in isolation; it sits within a system whose total cost can be many times the price of the chip itself.

This is one reason the semiconductor industry is such an important lens through which to understand the AI boom. The SIA reported particularly strong growth in logic and memory products in 2025. During the second quarter of 2026, worldwide chip sales increased another 35.1% from the previous quarter.

The growth is impressive, but investors should be careful about treating industry growth as synonymous with attractive returns for every company involved. Semiconductor markets are cyclical, capital intensive and highly competitive. A period of extraordinary demand can create strong incentives for suppliers to expand capacity, and if that capacity arrives faster than demand ultimately grows, pricing power can deteriorate surprisingly quickly.

The economics of the AI infrastructure cycle therefore depend not simply on how much computing is being deployed, but on who controls scarce parts of the supply chain and how durable those shortages prove to be.

Scarcity determines where the profits sit

The history of the semiconductor industry provides a useful framework for thinking about this. When supply is constrained and demand is strong, companies controlling scarce manufacturing capacity or differentiated intellectual property can capture significant economic rents. As capacity expands and competing products become available, those rents can move elsewhere.

Memory provides an obvious example. It is an essential component of AI infrastructure, particularly as increasingly complex models require large quantities of high-bandwidth memory. Yet memory is also a notoriously cyclical business, where changes in supply can have a dramatic effect on pricing and margins.

That makes the investment case quite different from owning a software company with a highly scalable business model. A semiconductor company can experience extraordinary earnings growth during a period of tight supply and then see those earnings normalise once the industry adds capacity.

The distinction between structural growth and cyclical earnings growth is consequently important. AI may create a long-term increase in semiconductor demand without every semiconductor company producing a permanently higher level of profitability. The investment opportunity depends on where the industry sits within that cycle, the company's competitive position and the valuation investors are already paying for future growth.

The second layer: infrastructure

The next layer of the AI economy is less glamorous but potentially just as important. Large-scale computing requires physical infrastructure, and that infrastructure requires electricity, land, cooling, networking and increasingly substantial grid investment.

The economics of AI therefore extend well beyond the technology sector. Data-centre construction creates demand for electrical equipment and construction services, while higher power consumption creates investment requirements across generation, transmission and distribution networks.

This physical constraint is becoming more important as the scale of AI deployment increases. A model can be improved through software, but a data centre still needs to be connected to the electricity grid and supplied with sufficient power and cooling capacity. In some locations, the availability of electricity may become a more important constraint on AI expansion than the availability of computing hardware.

That creates an interesting second-order investment theme. The companies that benefit from AI may not always have "AI" in their name or even operate in the technology sector. Some of the capital spending generated by AI could ultimately flow to businesses providing the infrastructure required to support it.

Investors should be cautious here as well. The existence of a structural infrastructure requirement does not automatically mean that every company operating in that ecosystem will generate attractive returns. Infrastructure projects can be competitive, capital intensive and subject to regulation, while the eventual utilisation rate of new capacity will determine whether the investment earns an adequate return.

The third layer: the cloud and model providers

Above the physical infrastructure sits the cloud and model layer, where the economics become more familiar to technology investors.

The major cloud platforms have the advantage of enormous distribution networks, established enterprise relationships and the ability to spread infrastructure costs across multiple services. Their position also allows them to capture value at several points in the AI stack, from infrastructure and compute to software services and enterprise applications.

But the scale of investment required is becoming an important part of the equation. The BIS estimates that the five largest hyperscalers were set to spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026, with some investment commitments running ahead of earnings and free cash flow.

That creates a fascinating investment tension. High capital expenditure can be evidence of strong demand and confidence in future growth, but it can also reduce near-term free cash flow and raise the amount of future revenue required to justify today's investment.

The IMF has highlighted a similar issue, noting that the returns on increasingly expensive and debt-financed AI investment may disappoint if adoption falls short of expectations. It also points to the possibility of interconnected exposures within the AI ecosystem, where a relatively small group of companies can simultaneously act as customers, suppliers, investors and financiers.

This does not make the investment cycle unsustainable by definition. It simply means that investors should pay attention to the relationship between capital expenditure and the revenue, margins and productivity improvements that eventually emerge from it.

The eventual prize is productivity

The most interesting part of the AI story may ultimately have little to do with chips or data centres. The larger economic prize is productivity.

If businesses can use AI to produce the same output with fewer labour hours, improve decision-making, accelerate software development, reduce customer-service costs or create entirely new products, the economic value of the technology can extend far beyond the companies selling the underlying infrastructure.

The IMF's 2026 analysis describes AI as potentially transformative for productivity, investment and labour markets, while its work on the UK highlights the opportunity for AI to address some of the country's persistent productivity challenges, particularly in high-skill, knowledge-intensive industries.

From an investment perspective, this creates a second phase of the AI story. The first phase is dominated by infrastructure spending: companies buying processors, building data centres and developing models. The second phase should be judged increasingly by what those investments actually produce.

That transition is important because it changes who captures the value. During an infrastructure boom, the suppliers of scarce computing capacity may enjoy exceptional economics. As AI becomes more widely available, the companies able to use it effectively could capture a larger share of the benefits through higher margins, faster growth or greater market share.

The difficulty is identifying those companies before the benefits are already reflected in their valuations.

AI does not eliminate the importance of valuation

There is a tendency during major technological shifts to assume that the size of the opportunity makes valuation secondary. History suggests otherwise.

A company can participate in an enormous structural growth market and still produce disappointing investment returns if the price paid for its shares already assumes most of the eventual success. Conversely, a business that looks less exciting may produce better returns if the market has underestimated its exposure to the underlying trend.

That is particularly relevant in AI because expectations have moved rapidly. Investors are not simply valuing current earnings; they are attempting to estimate what the industry might look like several years from now, how quickly AI adoption will spread, how much computing will be required and which companies will retain pricing power as competition develops.

The range of possible outcomes is therefore unusually wide. A modest change in assumptions around capital expenditure, model efficiency or enterprise adoption can have a substantial effect on the estimated value of businesses throughout the ecosystem.

There is also a counterintuitive possibility worth considering: improving AI efficiency could eventually reduce demand for some forms of computing even as AI usage increases. If models become dramatically more efficient, the amount of hardware required to produce a given unit of useful output could fall. Total demand may still increase, but the relationship between AI usage and hardware consumption would become less straightforward.

This is why the best AI investment analysis cannot simply extrapolate current semiconductor growth indefinitely. It needs to consider both sides of the equation: how much demand the technology creates and how rapidly the technology itself becomes more efficient.

Where the next phase could be decided

The AI investment cycle is still developing, and it is too early to know exactly where the largest long-term economic profits will settle. The evidence so far suggests that infrastructure providers have captured substantial value, while the next phase will depend increasingly on whether businesses outside the technology sector can translate AI capability into measurable improvements in productivity and profitability.

That makes the next few years particularly interesting. Semiconductor demand provides one measure of how quickly infrastructure is expanding, while hyperscaler capital expenditure indicates how much money is being committed to building capacity. Neither, on its own, tells us whether the investment is earning an adequate return.

The more useful evidence will increasingly come from the other side of the equation: revenue generated by AI products, reductions in operating costs, improvements in employee productivity and changes in competitive dynamics within individual industries. If those benefits begin to appear consistently, the investment case for AI becomes broader and potentially more durable than a technology infrastructure cycle alone.

For investors, that suggests keeping an eye on the entire chain rather than treating AI as a single trade. The companies selling the picks and shovels may continue to benefit, but the eventual economic winners could just as easily include businesses that use those tools exceptionally well. At the same time, the scale of capital being committed means that disappointing returns on infrastructure investment could create pressure in parts of the market even if AI itself proves to be genuinely transformative.

The central uncertainty is therefore not whether AI will change the economy. It almost certainly will. The harder investment question is how the value created by that change will be divided between the companies building the infrastructure, the platforms providing the technology and the businesses using it. That allocation of economic profit, rather than the technology itself, is likely to determine where the most durable investment opportunities emerge.