How to Manage the 4 Risks Hiding Inside Every AI-Heavy Portfolio
Huge amounts of money are being invested in AI infrastructure, but what happens if demand fails to live up to expectations? Savvy investors should review the AI concentration risk in their portfolios now.
Editor's note: This is the third article in a four-part series on AI concentration risk in growth portfolios. The first two are: Your Diversified ETF Isn't as Diversified as You Think, and AI Is a Supply Chain Rather Than an Industry: Here's Why. The final article will explain the corporate adoption timeline that determines which layers ultimately earn their valuations.
Every layer of the AI supply chain depends on one assumption holding: That the money being spent today gets paid back by revenue tomorrow. That assumption isn't guaranteed, and four distinct risks could break it.
Demand risk
The first risk is the simplest to state and the hardest to verify. Enterprises and consumers have to want AI capability enough to pay for it at the volume hyperscalers are underwriting. The revenue is real, but it is not yet close to matching the spending.
From just $107.88 $24.99 for Kiplinger Personal Finance
Become a smarter, better informed investor. Subscribe from just $107.88 $24.99, plus get up to 4 Special Issues
Sign up for Kiplinger’s Free Newsletters
Profit and prosper with the best of expert advice on investing, taxes, retirement, personal finance and more - straight to your e-mail.
Profit and prosper with the best of expert advice - straight to your e-mail.
Analysts estimate the annual gap between AI infrastructure spending and AI ecosystem revenue at roughly $600 billion, a figure that has widened rather than narrowed through 2026.
That gap doesn't prove demand will fail to show up. It proves that today's valuations are pricing in demand that has not arrived yet.
If it arrives on schedule, the companies building the infrastructure look prescient.
If it arrives two or three years late, a great deal of currently invested capital sits idle in the meantime, generating cost without matching revenue.
Capital spending risk
The second risk is that the spenders themselves change their minds. Microsoft, Amazon, Alphabet and Meta are projected to spend $700 billion to $725 billion combined in 2026, and every one of those companies answers to public shareholders who are already asking harder questions about the return on that spending.
Alphabet's stock fell roughly 7% in a single session in late July 2026 after the company raised its 2026 capital spending guidance, with investors specifically flagging concern over the return on that investment.
Capital spending plans are guidance, revised quarterly, not contracts. Every company in this chain can slow construction, delay chip orders or renegotiate power agreements if the return case weakens.
A slowdown doesn't require a crisis. It only requires one or two disappointing quarters of cloud revenue growth at the companies writing the checks.
Infrastructure bottleneck risk
The third risk runs in the opposite direction: Demand outrunning the physical capacity to meet it. Power, not chips, has become the binding constraint on the AI buildout.
Microsoft has disclosed a data center backlog worth roughly $80 billion that it can't yet fulfill because of power constraints. Separate estimates put the current U.S. data center power shortfall at more than 11 gigawatts today, growing toward nearly 50 gigawatts by 2028.
That constraint isn't just about equipment sitting in a factory somewhere. It's about the years-long wait to connect anything to the grid at all.
More than 2,060 gigawatts of generation and storage capacity are sitting in U.S. interconnection queues awaiting approval, according to Lawrence Berkeley National Laboratory, and the median project reaching completion in 2025 spent more than five years waiting in that queue before going live.
A company can have the capital, the chips and the building ready, and still be years away from switching anything on.
The equipment behind that power is just as constrained. GE Vernova, one of only a handful of companies in the world that build heavy-duty gas turbines at scale, reported an order backlog of 116 gigawatts in the second quarter of 2026, up from 100 gigawatts the quarter before, and it is already booking turbine reservations for 2031 delivery.
Writing a check today for new gas-fired power doesn't move a buyer to the front of a line that already stretches five years out.
A bottleneck sounds like good news for the companies that supply power and cooling equipment, and in the near term it often is. A persistent shortage also caps how fast revenue can grow across the entire chain, regardless of how much capital is available to spend.
Money can't buy electricity that doesn't exist yet, and permitting new energy takes years, not quarters.
Return-on-investment risk
The fourth risk is the one to watch most closely, because it determines whether the first three matter. Even if demand shows up, spending continues and bottlenecks clear, the infrastructure being built still has to generate a return that justifies its cost.
The divergence between AI capital expenditure growth and AI revenue growth is currently running around 46%, a figure that already exceeds the roughly 32% divergence observed during the telecom infrastructure buildout of the late 1990s, the cycle that preceded a multi-year correction in that sector.
That comparison is worth taking seriously without taking it too literally. The telecom buildout of that era did not turn out to be worthless; much of the fiber laid during that boom is still carrying traffic decades later.
It turned out to be built years ahead of the demand that eventually filled it, which meant a lot of capital invested at the peak was never recovered by the investors who supplied it, even though the infrastructure itself proved valuable.
AI's infrastructure could follow the same pattern: Durable and eventually useful, while still delivering a poor return for whoever financed the buildout at today's prices.
Three things to do about it
Look through every fund you own for exposure to the supply chain layers behind the AI buildout, not just the ticker you recognize. A fund can look diversified by holding count and still concentrate your money in a handful of companies whose revenue depends on the same four checkbooks.
Measure your overlap across funds, not within any single one. If your growth ETF, your S&P 500 index fund and a semiconductor fund all derive a meaningful share of returns from the same AI capital spending cycle, your effective concentration is higher than any single fund's fact sheet will show you.
Stress-test the shared assumption directly. Ask what happens to each holding, not the portfolio in the abstract, if hyperscaler capital spending growth slows from 60% to 15% next year. Some holdings would barely notice.
Others would take the full force of the correction. Knowing which is which before it happens is the entire point of this exercise.
The AI buildout may well justify itself over time. The risk worth pricing in today is not whether AI works. It's whether the bill comes due before the revenue arrives to pay it, and whether your portfolio is positioned to survive the gap either way.
Related Content
- Worried About an AI Bubble? 5 Ways to Ensure Your Portfolio is Prepared — Whether It Bursts or Not
- Is the 'AI Bubble' a Myth? Why Tech Experts Say AI's Boom Is Just the Beginning
- Is an AI Data Center Coming to Your Backyard? What It's Like in Retirement
- Will AI Pay Dividends for Your Firm? To Find Out, Budget for the Whole Iceberg, Not Just the Tip
- Using Google AI Tools Can Give Your Advisory Firm the Edge — If You Do These 5 Things First
