Is the AI trade a bubble, or is it just expensive?
The AI Big 10, Nvidia, Microsoft, Alphabet, Amazon, Meta, Apple, Tesla, Broadcom, Micron, and AMD, now make up 41% of the S&P 500. That is roughly the concentration tech and telecom held during the dot-com peak. Nvidia and Broadcom trade at price-to-sales ratios above 30, a level that has historically preceded sharp corrections. Deutsche Bank has called 2026 "1999 meets 1990," and still holds an S&P 500 year-end target of 8,000 at a P/E around 25x. The valuation concern is real, and Wall Street's own base case already assumes elevated multiples are permanent, not temporary.
What's different this time
Microsoft, Alphabet, Meta, and Amazon are funding their data centre build-out from operating cash flow, not debt or new equity. The dot-com bust was as much a financing collapse as a demand collapse, so self-funded spending does not carry the same failure mode. It can still be a poor capital allocation decision, but it will not trigger forced unwinding the same way. A National Bureau of Economic Research study from early 2026 found most firms report no measurable productivity impact from AI so far, despite executives projecting future gains. Well-financed spending can still fail to earn back its cost of capital.
The part that actually matters
The useful split isn't "bubble or not." It's structural demand versus narrative-driven capital. Data centre buildout tied to actual compute usage is one category. Capital flowing in because AI is the dominant story of the cycle is another. Both are priced into the same stocks right now, and they will not behave the same way when sentiment turns.
Earnings season ahead is what settles it. Current multiples assume the AI capital expenditure story gets validated by usage numbers. If the data says otherwise, the correction will not be gradual.
← All posts