Start with the raw spend. Google, Microsoft (MSFT), Amazon (AMZN), and Meta Platforms (META) put $168 billion of capital expenditure to work in the June quarter — up 74% year over year, according to Visible Alpha estimates. The four are expected to spend roughly $710 billion this year, and something in the neighborhood of $1 trillion in 2027. For scale: that annual figure rivals what the entire U.S. government spends on defense.
Meta is the most vivid case. The company has amassed roughly 20 gigawatts of computing capacity with another 14 gigawatts on the way — and Mark Zuckerberg has floated the idea that renting out the excess “is an option.” Read that again: one of the great cash machines in corporate history is buying so much compute that it is considering becoming a landlord for the surplus. The cost is real. Meta’s free cash flow is likely to be negative this year — the first time that has happened in its life as a public company.
While the buyers doubled down, the sellers got repriced. The PHLX Semiconductor Index fell 11% over two days, including a 4.65% drop Tuesday — down 599.62 points to 12300.52. SK Hynix lost 17% and Micron (MU) 15% over the same two days. Nvidia (NVDA), Broadcom (AVGO), AMD, and Intel (INTC) all fell. Even Caterpillar (CAT) — which sells the heavy iron and power equipment that data-center construction runs on — dropped 10% in two days. When the picks-and-shovels names and the earth-movers sell off together, the market isn’t questioning one company’s quarter. It’s questioning the spending war itself.
Here is the part most investors haven’t internalized: chip stocks now make up roughly 18% of the S&P 500’s market value, versus about 5% five years ago. And of the four hyperscalers writing those capex checks, only Alphabet (GOOGL) has actually beaten the S&P 500 this year. The spenders are lagging the index they dominate, and the index has quietly become a semiconductor fund with a stock market attached.
If you want a feel for how hot the arms race is burning, look at what the model companies are giving away. OpenAI and Anthropic are now handing Y Combinator startups $500,000 credit packages — Anthropic’s is up from $30,000, and OpenAI’s $500,000 in free credits comes with an optional $1.5 million more in exchange for equity. Some startups have collected more than $3 million in credits — roughly the size of a median U.S. seed round, per PitchBook. Y Combinator companies could see up to $800 million in combined credits next year. And in the category of numbers that shouldn’t be possible: a compute deal between xAI and Anthropic runs $1.25 billion a month.
Giveaways at that scale are what companies do when market share today matters more than margins today. It worked for cloud computing. It also worked, for a while, for 1999 telecom. The honest answer is nobody knows which one this is — which is exactly why sizing, not conviction, is the discipline that matters.
If you hold an S&P 500 fund, roughly 18 cents of every dollar is now riding on semiconductors — nearly four times the weight of five years ago. You didn’t choose that allocation; the market drifted you into it. There’s nothing wrong with owning the AI build-out. There is something wrong with owning it accidentally, at maximum size, without a counterweight.
First, know what you own. Pull up your index fund’s top holdings and add up the chip and hyperscaler weight — the number will surprise you. Second, barbell it. We pair the growth concentration with things that get paid regardless of who wins the spending war: dividend payers and utilities on one side, and a short-Treasury sleeve still paying 4%-plus on the other. Cash flow now is the antidote to capex faith. Third, own the picks-and-shovels on purpose, at a size you chose. Our themed sleeves hold the AI build-out deliberately — sized so that a two-day, 11% semiconductor drawdown is a headline, not a household event. That’s the difference between riding a boom and being ridden by one.