Ask five of Wall Street’s most legendary investors “is AI a bubble?” this year and you’d get five different portfolios, not five different opinions. In the first quarter of 2026, Warren Buffett’s Berkshire Hathaway more than tripled its stake in Alphabet. In the same quarter, Bill Ackman sold 95% of his Alphabet position and rotated into Microsoft. Stanley Druckenmiller sold his entire Alphabet stake too, cut his portfolio from 62 stocks down to 22 — and bought AI chipmaker Broadcom. Cathie Wood’s ARK funds sold Nvidia, then bought it back two months later. Michael Burry, the investor who called the 2008 housing crash, bought put options betting against both Nvidia and Palantir. Same industry, same information, five completely different bets. That’s not confusion — it’s five investors choosing where to stand in a four-story building, and the ground floor carries a risk called stranded cost that has nothing to do with whether AI ever actually works.
Picture the AI industry as a building with four floors. The first floor is power and physical infrastructure: power plants, transformers, cooling systems, land. The second floor is chips — GPUs, custom silicon, the memory that feeds them. The third floor is cloud platforms like Microsoft, Amazon, and Google, which spend heavily on AI while also selling AI services. The fourth floor is applications — companies promising to build something useful with all of it, with the least proof so far. The floor an investor picks says more about their read on AI than any “bullish or bearish” headline ever could. And the first floor has a feature the other three don’t: it gets built whether AI succeeds or not.
That’s not a metaphor — it’s happening in the U.S. power grid right now, in a very literal, billable way. Data center developers have flooded power utilities with requests for new grid connections, and because reserving a spot in line costs them almost nothing, they file the same project with multiple utilities simultaneously and pick whichever offers power first — inflating the queue far beyond real demand. Utilities have started responding: PJM, the largest U.S. grid operator, saw its capacity auction prices jump nearly tenfold in its most recent round as it tries to pay for supply to meet inflated demand forecasts. Virginia utilities have proposed new rate structures that force data centers to pay for 60–80% of their contracted power whether they use it or not, specifically to stop the cost of overbuilt infrastructure from landing on ordinary ratepayers. The pattern is the same everywhere: infrastructure gets reserved and built based on demand that may never fully show up, and someone has to pay for the gap between “requested” and “used.”
There’s a term for this in utility economics: stranded cost. It originally described what happens when a power plant or grid investment can’t earn back its cost — usually because demand didn’t materialize the way it was projected to — and regulators let utilities recover that loss through rates paid by everyone, not just the customer who caused it. The AI buildout is reproducing that exact mechanism at a much larger scale. Whether or not AI ever delivers on its promises, the power plants, transformers, and land already committed to serving it don’t disappear. Someone pays for them either way.
This is precisely why the four-floor framework matters for figuring out who’s exposed. First-floor and second-floor investments are locked in regardless of AI’s ultimate success — the money is already spent, and stranded cost risk concentrates there. Third-floor companies have an escape hatch that Amazon proved works in 2001: after the dot-com crash wiped out 90% of its stock price, Amazon survived because it could simply cut spending — a seller of infrastructure loses revenue when demand drops, but a buyer of infrastructure just stops buying. Cisco, by contrast, was a second-floor company selling the routers everyone needed for the internet buildout; it was right about the demand and still took 25 years to reclaim its 2000 stock price, because the market had already priced in 20 years of growth within 2. Being right about the technology and being right about the timing turned out to be two completely different bets.
None of this means “AI is a bubble” in the simple sense — the difference this cycle is that the companies funding the buildout (Alphabet, Microsoft, Amazon) generate enormous free cash flow and are mostly self-funding it with equity, not debt. A debt-funded bubble collapses in one snap; an equity-funded overbuild just grinds through several years of weaker returns. But “AI overall isn’t a bubble” and “the first floor is carrying stranded cost risk regardless of outcome” are two different claims — and Wall Street’s biggest names are positioning as if only the second one is true.
My take: The bill for infrastructure that never gets used still comes to your door.
Not advice. Just how I see it.