Opinion · July 2026
The AI Trade Split in Two on the Same Day: Hardware Up, Software Down
Wednesday’s index moves were nothing — the S&P 500 fell 0.12%, the kind of session that looks like a rounding error in a year-end chart. Underneath it, the AI trade split cleanly into two pieces that moved in opposite directions on the same news. Super Micro Computer jumped 19.84% on guidance for 2026 gross margins to roughly double. Broadcom, Talen Energy, Vistra, and CoreWeave — chips and the power to run them — all rose too. At the same time, ServiceNow fell 6.47%, Datadog fell 3.54%, Palantir fell 6.10%, and Coinbase fell 5.53%, on the same day a Goldman Sachs report on $489 billion in AI-related debt issuance made the rounds.
One trade, two risks
This isn’t two unrelated stories. It’s the same AI-capex trade being priced from its two ends at once. The hardware and power names are the supply side of the buildout — and Wednesday's news (Super Micro's margin guidance, continued strength across chips) was, on its face, confirmation that demand for that supply is real and still growing. The enterprise-software names that fell are the demand side of a different, adjacent question: with $489 billion in AI-related debt now on balance sheets across the industry, is that spending increasingly financed with leverage rather than organic cash flow, and if so, which companies are exposed to that unwind first? High-multiple software names with growth priced in for years out are exactly the kind of asset that gets repriced first when the market starts asking that question, even when nothing about their own quarter has changed.
Where this connects to my own thesis
This is the same fork I raised on the AI Capex Reality Check and in the Physical Limits of Compute whitepaper: real, growing demand for AI infrastructure doesn’t automatically mean every AI-adjacent stock is mispriced in the same direction. Wednesday is a small, live example — the hardware side got a demand-confirming data point and rallied, while the debt-financing overhang landed hardest on software names whose valuations already assume years of growth. Both reactions can be correct at the same time, about different parts of the same buildout.