AI’s capex surge tests investor discipline: ‘I’m more interested in picks and shovels than the gold mine’
institutional investors question whether AI-driven investment returns reflect durable growth (Unsplash/Mohamed Nohassi)
As market concentration deepens and infrastructure spending accelerates, institutional investors question whether today’s AI-driven investment returns reflect durable growth or the later stages of a crowded trade.
The acceleration of capital into AI has become one of the defining forces in global markets, with AI-linked companies accounting for an outsized share of equity returns.
A single company — Nvidia — accounted for roughly 15 per cent of the S&P 500's total return in 2025, underscoring the disproportionate role a small number of AI-linked stocks have played in driving index performance.
According to S&P Dow Jones Indices data compiled by industry analysts, cited by Statista, the Magnificent Seven tech companies, including Microsoft, Alphabet and Nvidia, accounted for 42 per cent of the S&P 500's 17.9 per cent return last year.
For institutional investors, the debate is less about whether AI will reshape the economy and more about what today's concentration and valuation profile imply.
Investor unease over the scale of AI-related capital expenditure is rising, with both equity and credit markets reacting to concerns that spending may be overshooting sustainable demand. Chief investment officers of public pensions increasingly warn that positioning risk is rising, even among allocators not explicitly targeting AI exposure.
"We tend not to make portfolio changes based on near-term trends, and in that regard the valuation of AI is no exception," says Bill Atwood, executive director of the $11bn Illinois Firefighters' Pension Investment Fund. "With that in mind, I do pay attention to the growth of AI, the associated need for capital and the implications for our portfolio."
That emphasis on discipline is echoed at the $36.8bn Police and Firemen's Retirement System of New Jersey, where executive director Greg Petzold says the fund is avoiding reactive shifts in response to AI enthusiasm. Rather than attempting to judge whether current valuations fully reflect long-term earnings potential, PFRSNJ is relying on its investment policy and its rebalancing framework to manage exposure.
"There's no question there's a lot of excitement, and quite a lot of speculation, around AI right now," says Petzold. "Our job isn't to predict where that excitement goes next, but to stay disciplined. Sticking to our policy allows us to participate if the fundamentals are real, while limiting the risk if expectations get ahead of reality."
Hyperscaler spending probes capacity and demand
Valuation concerns remain particularly acute across parts of the AI supply chain.
Semiconductor and AI-hardware manufacturers continue to trade at elevated multiples relative to historical norms, as investor expectations price in sustained demand growth. PwC notes in its 2026 global semiconductor industry outlook report that the semiconductor market is expected to exceed $1tn by 2030.
At the same time, capital expenditure by hyperscalers — the largest cloud and internet infrastructure companies such as Amazon, Microsoft, Google Cloud and Meta — has accelerated sharply as they scale infrastructure for AI workloads.
According to industry forecasts, these four companies were expected to spend more than $345bn in 2025, a roughly 50 per cent year-on-year increase, with much of that investment directed towards AI-optimised data centres, networking and power infrastructure.
Capital spending by AI hyperscalers could exceed $500bn in 2026, with current consensus around $527bn, according to Goldman Sachs, even as the pace of growth begins to moderate. Analysts at Goldman say forecasts have repeatedly underestimated actual outlays, suggesting further upside as companies continue investing heavily in data centres, chips and power infrastructure.
Industry experts also flag emerging constraints around power availability and grid infrastructure, alongside longer-term concerns that rapid deployment of graphics processing units, the specialised chips used to train and run AI models, could ultimately outstrip end-user demand. Meta, for example, is pursuing new energy deals — including nuclear power agreements — to secure electricity supply for AI expansion.
But Atwood says the accelerated expansion of industrial computing has begun to compress returns in certain segments.
"The need for, and investment in, industrial computing continues to expand, but the return premia in those investments seem to have compressed, and there is emerging discussion of possible oversupply," he says. At the same time, he notes that barriers to entry are rising elsewhere in the ecosystem, particularly as regulatory scrutiny intensifies around data centres and access to physical infrastructure.
That divergence is increasingly impacting the way large asset owners position themselves. While headline AI beneficiaries dominate public-market exposure, some allocators see more resilient economics further downstream. Both Atwood and Petzold point to energy and infrastructure as areas where AI demand intersects with longer-term structural needs.
"Current capacity for generation and distribution of electricity can barely meet existing demand, and much of our infrastructure is operating beyond its design life," says Atwood.
"There is no end in sight to AI's expanding power requirements. As a result, I am fairly confident those able to provide the capital to upgrade and expand the power grid will be sufficiently compensated."
Petzold says PFRSNJ is similarly wary of concentration risk in public markets, and is instead seeking exposure through private investments that support the broader AI ecosystem without relying on continued outperformance by a narrow group of mega-cap stocks.
Areas of interest include power generation, transmission and cooling technologies — segments that benefit from AI growth but retain standalone economic value.
In public equities, concentration has made portfolio construction more challenging.
Fund managers warn that crowded positioning in the same group of AI-linked stocks has increased correlations and reduced diversification benefits.
In response, some institutional investors are looking further along the value chain, including utilities exposed to rising power demand and industrial companies supplying grid and energy infrastructure.
In private markets, competition for AI-themed deals has intensified, with valuations in parts of the early-stage and growth universe approaching levels last seen during the 2021 investment boom. Atwood says his focus remains on indirect exposure rather than backing AI companies outright.
"I am less interested in AI in terms of direct investments, and more as a driver of ancillary opportunities," he says, pointing to rising capital needs across fibre networks, cellular infrastructure, satellite transmission and broader connectivity.
"Using a tired metaphor," Atwood added, "I'm more interested in the picks and shovels than the gold mine."
While the long-term economic potential of AI remains substantial, institutional investors say the current environment calls for valuation discipline, diversification within the ecosystem and caution around consensus exposure.
For asset owners and asset managers, the AI challenge is to avoid assumptions about today's leaders, while identifying mispriced enablers before the market takes on the contours of a bubble.
"Those of us who have lived through the global financial crisis and dotcom bubble certainly see similarities," says Petzold. "We try to tune out the day-to-day noise, keep a long-term mindset, and make sure the portfolio is positioned to weather whatever the next cycle brings."