The AI Buildout: Exposure, Uplift and What It Means for Diversification
Aureus INVESTMENT Perspectives | OCTOBER 2026
If you are an avid or even casual reader of the news, it has been impossible to avoid the topic of AI over the past several years. Popular angles have included best-positioned companies, circular financing arrangements, the risks to white-collar jobs, the consumer impact from higher electricity prices, broad concerns around data center construction, and more recently, existential threats to humanity.
In this Investment Perspectives, we explore just how pervasive and influential AI has become in the world of investing, and most importantly, how it is influencing our conversations around stocks, investable asset classes, and the economy.
Our conclusion is that the AI buildout has put a single driver behind nearly every asset class we own, and that diversification today requires carefully monitoring thematic exposure, not just asset-class weights.
Estimates of AI’s economic impact range widely because they track different targets: the capital being spent, the market value created, and the productivity gained. The first two are large and measurable today; the third is still mostly a forecast. This piece follows the capital: where it is being spent, where it lands, what it adds to growth, and what it is costing the rest of the economy.
PUBLIC EQUITIES: SPENDERS, SUPPLIERS, AND CAPEX EXPOSURE
By JPMorgan’s count in early October, 41 AI-linked companies made up 49% of the S&P 500 Index and drove roughly two-thirds of its gains this year. Broader definitions of what counts as an AI company push that share past 60% of the index. Behind those numbers is one input: the five largest hyperscalers are on track to spend roughly $800 billion on capital expenditures in 2026, up from about $450 billion in 2025, with 2027 estimates now clustering around $1.2 to 1.3 trillion. Analysts have revised 2026 and 2027 estimates up by roughly twothirds since the start of the year, so any figure here is likely more a floor than a forecast. That spending drives every layer of the AI infrastructure buildout:
| Layer | What the capex buys | Representative public exposure |
| Spenders | The buildout itself Exposure is to the value/return on that spend | Microsoft, Alphabet, Amazon, Meta, Oracle, CoreWeave |
| Compute | GPUs, custom accelerators, CPUs ~30% of hyperscaler capex flows to Nvidia alone | Nvidia, AMD, Broadcom, Marvell |
| Memory and foundry | HBM, DRAM, leading-edge fabrication This is the tightest supply chain in the stack | SK Hynix, Samsung, Micron, TSMC |
| Networking and optics | Switches, optical transceivers, fiber, interconnect | Arista, Cisco, Coherent, Lumentum, Corning, Credo |
| Equipment and tools | Lithography, deposition, test | ASML, Applied Materials, Lam, KLA |
| Servers and racks | Assembly, liquid cooling, power distribution | Dell, Super Micro, Vertiv, Eaton, Schneider |
| Physical plant | Construction, HVAC, electrical contracting, transformers, switchgear | Quanta, EMCOR, Comfort Systems, Trane, Hubbell, GE Vernova |
| Power | Generation, nuclear restarts, gas turbines, behind-the-meter supply | Constellation, Vistra, NRG, Talen, GE Vernova, Caterpillar, Bloom |
| Commodities | Copper, aluminum, uranium, natural gas, silver, helium | Freeport, Southern Copper, Cameco, EQT |
| Capital raisers | Underwriting, M&A advice, structured and private credit origination | Goldman, Morgan Stanley, JPMorgan, Blackstone, Apollo, Blue Owl |
Names are illustrative of each layer, not recommendations.
The table also shows that the spenders and the suppliers carry different risks. Suppliers are paid as the capex is spent, whether or not it ever earns a return, so they bear the cyclical risk: if spending pauses, their demand can stop almost immediately. The hyperscalers carry the risk that the spending never pays off. They only win if the capex generates a satisfactory return, but if it does, that benefit accrues for many years.
BEYOND STOCKS: WHAT MOVES EQUITIES NOW MOVES EVERYTHING
Real estate. Data centers are the most obvious exposure, through both public REITs and privately held assets, and that exposure has been growing steadily in core real estate funds. Less obviously, certain residential markets, San Francisco above all, are seeing outsized demand from newly wealthy AI employees.
Infrastructure. The binding constraint has moved from chips to power and grid equipment: gas turbines, transformers and switchgear, transmission, nuclear restarts, and on-site generation built to bypass interconnection queues. Infrastructure funds own the pipes, the generation, and increasingly the campuses themselves, making any infrastructure holding a likely AI holding.
Credit. This is where the exposure is newest and least visible. JPMorgan projects about $300 billion a year of AI and data center debt issuance for the next five years, much of it in off-balance-sheet structures such as Meta’s $27 billion Hyperion joint venture with Blue Owl, where the debt sits with the private credit partner rather than on Meta’s books. A client holding an investment-grade bond fund, a private credit fund, and an insurance-backed annuity owns the AI buildout three times over.
Private equity. Private markets are where AI exposure is largest relative to what clients expect. On the venture side, AI megadeals produced the two largest quarters of venture investment on record in the first half of 2026, and the ten largest deals in the second quarter alone totaled $105 billion, with a handful of foundation-model companies such as OpenAI and Anthropic taking a large share. On the buyout side, the large sponsors have spent the past three years acquiring the picks and shovels: data center developers, electrical and mechanical contractors, power-equipment makers and the land and interconnection rights beneath the campuses. Because private marks lag public prices by a quarter or more, this exposure also shows up late. A client whose venture and buyout funds appear uncorrelated to their equity portfolio may simply be looking at last quarter’s AI trade.
International. Korea is the clearest case, in both directions. The Kospi rose 76% in 2025 and more than doubled again to a record above 9,000 in June of this year, then fell nearly 40% in the four months since. It is still up roughly 30% for the year. SK Hynix and Samsung, which each passed $1 trillion in market value in May and together make up around 40% of the index, drove both legs. Taiwan, Japan and the Netherlands follow the same pattern. Buying international equities is not a clear path to diversifying away from AI, and Korea’s drawdown is a cautionary example of what that concentration looks like on the way down: the index moved 3% in a single session recently on nothing more than a US AI company filing to go public.
AI AND US GDP GROWTH: WHAT IT ADDS AND WHAT IT COSTS
AI investment is contributing meaningfully to US real GDP growth in 2026. Depending on definition, estimates range from one-third to one-half of GDP growth coming from AI investment. Without it, the economy would be growing at around 1%.
Two features of this growth matter for the forward view.
It is investment-led, not productivity-led. The Fed’s Beige Book says AI has not yet significantly affected staffing, and a measurable productivity dividend is still not observable in the data.
It has cannibalized other investments. Non-residential business investment outside tech fell year-on-year for six straight quarters through the first quarter of 2026, and data centers and power generation are the only construction categories growing. With the 10-year Treasury yield at around 5%, the highest in nearly two decades, the capital the buildout is absorbing is a real cost to the rate-sensitive parts of the economy, housing most of all.
Over the next year or two, it comes down to one thing: whether hyperscalers keep spending. If hyperscaler spending rises to $1.2 to 1.3 trillion in 2027, as consensus now has it, the contribution to GDP growth holds or rises. A flattening of capex in 2028, which several analysts expect as spending catches up with revenue, would take away a third of the economy’s growth within a year unless a productivity dividend has arrived to replace it. That hand-off, from investment-led to productivity-led growth, is the single most important macro question of 2027 and 2028.
WHY DIVERSIFICATION NOW MEANS MONITORING A THEME, NOT AN ASSET CLASS
A portfolio can be diversified by every traditional label and still be a single bet on the AI capex cycle continuing. Consider a client with a US large-cap index fund, an international equity fund, a core bond fund, a private credit allocation, an infrastructure fund, a real estate sleeve, PE and VC investments, and a home in the Bay Area. By asset class that has eight exposures. By theme it is one: the index is dominated by the spenders and the chip layer; the international fund’s return has come from Samsung, SK Hynix, TSMC and ASML; the bond fund holds hyperscaler and data center paper; the private credit fund owns the campus joint ventures; the infrastructure fund owns the turbines and the grid; the real estate sleeve owns the data centers; the venture sleeve is marked to the AI labs’ last round; and the house is priced by AI liquidity events. A pause in capex or a hyperscaler cutting guidance would move all eight in the same direction on the same day.
The traditional diversification framework was built for a world where asset classes responded to different drivers: equities to earnings, bonds to rates and inflation, real estate to local supply, infrastructure to regulated returns. The buildout has put one driver behind all of them, and their returns have become correlated as a result.
WHAT THIS IMPLIES FOR HOW WE MANAGE PORTFOLIOS
None of this is a reason to abandon the theme. It has been the right place to be, and our clients have benefited from it. Rather, it is a reason to know exactly how much of it you own:
Measure thematic exposure across the whole balance sheet, including private funds and real estate, and report it alongside asset-class weights.
Distinguish between the three types of AI exposure: suppliers, spenders, and lenders, which carry different risks and should not be netted into one number.
Remember that the holdings that have lagged are the assets that would benefit from a capex slowdown. Their underperformance is the cost of the diversification they provide.
Watch for the hand-off: hyperscaler spending relative to revenue, memory pricing, spreads on data center debt, and the first signs of measured productivity gains. These, rather than index levels, will tell us whether investment-led growth is becoming productivity-led.
The AI buildout is the most concentrated capital cycle since the 1990s telecom build. The uplift is real, measurable, and still rising. So is concentration. The purpose of diversification has always been to avoid overexposure to a single variable, and today that means measuring it by theme as well as by asset class.
This commentary reflects the opinions of Aureus Asset Management as of the date of publication and is subject to change without notice. It does not constitute investment, legal, or tax advice, nor a recommendation to buy or sell any security. Statistics are drawn from sources believed to be reliable, including JPMorgan, ING, Federal Reserve, BEA, S&P Global, Goldman Sachs, KPMG, and Bloomberg, but accuracy is not guaranteed. Past performance is not indicative of future results.
