The scale of the AI infrastructure buildout
The artificial intelligence race has moved from software to steel and silicon. Cloud giants are now spending sums that rival national budgets to house the servers powering chatbots, image generators and enterprise AI tools, and that capital splurge is reshaping several corners of the stock market.
John Mowrey, chief investment officer at NFJ Investment Group, estimates that the largest cloud providers, known as hyperscalers, are on track to spend between $750 billion and $800 billion a year. Some projections push the figure as high as $1 trillion, a level that would equal 2.5% to 3% of US gross domestic product.
"Hyperscalers are going to spend maybe between $750 and $800 billion [a year]. Some forecasts even have it up to a trillion, and that would put it at 2.5 to 3% of U.S. GDP, which is just extraordinary for a capital market," Mowrey said.
Much of that money is flowing into data centres, the climate-controlled warehouses of computing power that form the physical layer of the AI boom. For investors who want exposure to the trend but worry that marquee names such as Nvidia have already run too far, the buildout offers a range of so-called pick-and-shovel opportunities across four areas: semiconductors, real estate, energy and cooling.
Semiconductors: the equipment makers behind the chips
Every data centre needs processors, and demand for the specialised chips that run AI workloads has outpaced the industry's ability to manufacture them. Craig Ellis, research director and senior semiconductor analyst at B. Riley Securities, describes semiconductors as one of the most acute bottlenecks in the entire buildout.
"We're at a point where undersupply is so severe that there needs to be a multi-year period of unusually strong capex growth in front of us, and that capex growth is something that is very investable because it has the potential to continue to lift expectations for revenues and earnings," Ellis said.
Instead of chasing the best-known chip designers, Ellis points to the companies that sell the machinery used to fabricate semiconductors. Applied Materials, the world's largest semiconductor equipment company, is a prime example. Its Semiconductor Systems division accounts for roughly 73% of revenue, according to its most recent annual filing, and supplies tools used in the production of advanced computing, logic and memory chips. Because nearly every leading chip and display passes through its equipment at some stage, the company is exposed to the whole chipmaking ecosystem rather than a single product cycle.
Ellis also highlights Lam Research, which makes equipment for memory and storage chips. Its portfolio is narrower than Applied Materials', but B. Riley Securities believes it is particularly well placed to benefit from a surge in new capacity investment over the next two years. The broker has raised its earnings estimates for Lam by 25% to reflect that outlook.
Marvell Technology is another candidate, especially in the networking segment that connects thousands of chips inside a data centre so they function as one machine. The company commands a market value of roughly $195 billion, according to analytics firm FactSet, putting it in the same tier as Nvidia and AMD. It has also drawn a direct investment from Nvidia as part of a partnership on next-generation networking technology.
Marvell's largest customer is currently Amazon Web Services, a relationship that has been lucrative but also leaves the company heavily tied to one client. Ellis argues that this concentration is an opportunity: if Marvell can win more hyperscaler business and scale its networking products, its profitability could rise meaningfully.
Chip stocks are not without risk. The Philadelphia Semiconductor Index, which tracks AI-linked chip stocks, fell 26% during a recent bout of anxiety that the spending boom might not pay off. Ellis contends that the sell-off already priced in much of that fear, and the numbers have since moved in his favour. Since the low on 28 July, Applied Materials, Lam Research and Marvell have each rebounded by roughly 22% to 30%.
Real estate: landlords of the digital economy
Chips need buildings, and real estate investment trusts, or REITs, own many of the specialised facilities that house them. These trusts rent out climate-controlled space with backup generators and high-speed connectivity, offering investors a way to own the physical backbone of the AI boom.
Patrick Wilson, a portfolio manager on the real estate securities team at CenterSquare Investment Management, says the investment case has strengthened as the industry shifts from training AI models to running them, a stage known as inference. Training can happen in large rural campuses built for cheap land and power, but inference works best in facilities close to major population centres, where shorter distances cut the delay between a user request and a model response. Established REITs already control the scarce, carrier-dense urban real estate that new entrants would struggle to replicate.
REITs also offer a tax advantage for ordinary investors.
"REITs are only taxed once, provided that they satisfy IRS tax rules. So, if they pay out 90% of their taxable net income into a dividend, there's no corporate level tax that they pay," Wilson explained.
Wilson points to Equinix and Digital Realty as the two largest ways to play the theme. Equinix operates data centres that host servers for thousands of companies, from Fortune 500 firms to major cloud providers such as AWS and Google Cloud. Digital Realty has spent more than two decades acquiring and building similar facilities, and Wilson notes that demand for secure, climate-controlled server space still exceeds supply. That imbalance has allowed Digital Realty to push rents higher, while its broad tenant base makes it more resilient than newer competitors dependent on a handful of large contracts.
The risks are real. Data centre REITs are sensitive to rising interest rates, which lift borrowing costs and weigh on property valuations. They also rely on premium rents, so a wave of new supply could squeeze margins. Geography is another constraint: while proximity to cities reduces latency, suitable land is scarce. "They need a lot of land, and that's not necessarily found in midtown Manhattan," Wilson said.
Still, Wilson regards established REITs as a steadier bet than newer AI infrastructure firms that are carrying heavy debt. If demand slows or short-term customers leave, those leveraged operators remain on the hook for long-term bills and interest payments.
Energy: powering the machine
AI data centres consume enormous amounts of electricity, and that appetite has transformed the outlook for utilities. Andrew Bischof, utilities analyst at Morningstar, notes that US electricity demand grew at just 0% to 0.5% a year for decades. The AI buildout has lifted that forecast and turned slow-growth yield stocks into vehicles for capital appreciation.
"You're now seeing more growth-oriented investors coming to utilities because they can provide that six to eight and sometimes 10% annualized growth over the five-year forecast," Bischof said.
American Electric Power, an Ohio-based utility holding company, is one of the most direct beneficiaries. It distributes electricity to more than five million customers across 11 states and plans to invest $78 billion between 2026 and 2030 to build the infrastructure needed to serve data centre-driven demand. Morningstar expects that spending to translate into 9% average annual earnings growth through 2030, a strong pace for a regulated utility.
Not every utility is equally attractive. Bischof cautions that in some regions, particularly the Mid-Atlantic, the rush to serve data centres has pushed power prices higher and raised concerns that retail customers will bear the cost. He prefers utilities in areas where regulators and communities are aligned behind the boom, naming DTE Energy, Alliant Energy and Evergy as examples.
For a more contrarian energy play, David Trainer, chief executive of investment research firm New Constructs, argues that renewable sources are growing but cannot yet deliver the baseload power that massive AI facilities require. That leaves fossil fuels as a practical necessity, and Trainer sees traditional energy stocks as a rare chance to buy infrastructure exposure at value prices. He points to refiners Valero and HF Sinclair, which he believes the market is pricing as if their profits are destined for a permanent 30% to 40% decline.
Cooling: keeping the heat at bay
All that electricity generates heat, and standard air conditioning cannot keep up in the densest AI facilities. Nick Lieb, industrials analyst at Morningstar, says cooling and power equipment offer a different kind of AI trade. While flagship GPUs are refreshed every year or two, the core technology behind data centre infrastructure evolves far more slowly, giving the sector a steadier profile.
Vertiv is Lieb's leading example. The company owns the Liebert brand, which pioneered the original computer room air handling unit in the 1960s.
"Vertiv owns the original data center cooling brand called Liebert, which actually invented the original computer room air handling unit… in the 1960s. That base technology is still quite relevant today [and] sells a lot of computer room air handling units," Lieb explained.
The stability comes with concentration risk: more than 80% of Vertiv's revenue comes directly from the data centre market, so any slowdown in hyperscaler spending would hit the company hard.
Investors who want more limited exposure can look at Eaton, a power management company for which data centres account for only about a quarter of sales. The rest comes from commercial buildings and other parts of the electrical grid. That diversification provides a buffer if data centre growth slows, and Lieb adds that Eaton is well placed to benefit from a US electrical grid that he describes as old and long overdue for replacement. The company produces transformers and switchgear, both of which are needed for new data centres and for the broader grid upgrade.
What could go wrong
The buildout is not without doubters. Lieb worries that the sheer scale of spending could become a long-term risk, noting that hyperscalers have reached a point where the cost of new data centres is starting to exceed the cash they generate. That gap is forcing tech giants to raise more debt and issue equity to fund their ambitions.
Trainer calls the current market a "crowded trade" and warns that it is hard to find an obvious stock that is not already expensive. He argues that large firms such as Amazon and Microsoft, which have been publicising their heavy capital expenditures, may not be able to sustain that spending pace. It is a race, he suggests, that some current participants lack the balance sheets to finish.
The long view
Yet the size of the opportunity is difficult to dismiss. Mowrey argues that the data centre boom is not merely a capital markets story but the physical foundation for a technology shift that could spread through much of the economy.
"AI adoption is still in its early innings… I think we're just scratching the surface," he said. "The broad enterprise application across healthcare, manufacturing, and financials, that's going to take years to filter in."
For investors, the question is not whether AI infrastructure will be built, but which parts of the supply chain can turn that construction into durable returns.