Over Our Skis
I've swung more cautious. Long term, I still believe AI is the most important technology of my lifetime. Short term, I think we've gotten over our skis.
The worry is the financing. The build-out is being paid for with enormous amounts of capital, much of it borrowed, and those costs come due on a schedule. AI revenue is growing fast, but not fast enough to cover them yet.
The napkin math. The largest cloud companies will spend about $800 billion on data centers and equipment this year, and about $1 trillion a year through 2030 (Vanguard). Each year's spending gets written off over the following six years, and the borrowing behind it carries interest, so the yearly bill climbs fast.[1] AI revenue is about $175 billion today: OpenAI near $50 billion, Anthropic above $65 billion, Microsoft near $37 billion, and the rest. After power, people and research, only about half of it is left to pay that bill.[2] Even if revenue doubles next year and keeps growing quickly after that, the shortfall widens every year:
The build-out’s yearly bill vs. what AI revenue can pay toward it, in billions of dollars. Usable is the half of AI revenue left after operating costs. Napkin math, rounded.
| 2026 | 2027 | 2028 | 2029 | 2030 | |
|---|---|---|---|---|---|
| Depreciation (forecast) | 135 | 300 | 465 | 635 | 800 |
| Interest (forecast) | 30 | 65 | 95 | 125 | 160 |
| Total cost | 165 | 365 | 560 | 760 | 960 |
| AI revenue (generous case) | 175 | 350 | 600 | 900 | 1,200 |
| Usable (half) | 90 | 175 | 300 | 450 | 600 |
| Gap | −75 | −190 | −260 | −310 | −360 |
By 2030, covering the bill would take about $1.9 trillion of AI revenue, not $1.2 trillion.
So I've pulled in. Cash is now 28.83% of the portfolio, up from 19.71% a week ago. I moved my index position from the Nasdaq-100 (QQQ) to the broader S&P 500 (SPY), and I sold Coherent and my SpaceX call options. I'm keeping my core holdings: TSMC, AMD, Rocket Lab, gold and bitcoin.
This isn't a call that AI is a bubble about to pop for good. It's a call that the build-out has run ahead of the cash flows. If revenue catches up faster than I expect, I'll be wrong, not early, and my returns might be muted. If it doesn't, I'd rather have cash ready to buy great companies when they're cheaper.
[1] Spending of $800 billion in 2026 and $1 trillion a year after that, written off evenly over six years, the top of the cloud companies' five-to-six-year range. Interest on $450 billion of new AI borrowing a year, at 7%. The depreciation and interest rows are my own arithmetic on those estimates.
[2] Today's revenue counts the labs plus about $50 billion I estimate for AWS and Google, less about $25 billion the labs pay the cloud companies. I assumed it doubles in 2027, then grows 70%, 50% and 35% a year, and that half is left after operating costs. That path is well above published forecasts, most of which put total AI revenue near $130 billion to $240 billion by 2030.