Semiconductors doubled in the first half of 2026 and gave a large chunk of it back in July. The story attached to that move is a rotation: sell the chips, buy the hyperscalers. I think the story is roughly backwards, and it's worth explaining why.
What actually happened
| Period | Move |
|---|---|
| H1 2026, chip index | roughly doubled |
| July 2026, chip index | ~−18% |
| July 23, megacaps | worst day since April 2025 |
| Alphabet / Amazon, Jul 23 | −5-6% / −4% |
The problem with the rotation thesis
If you sell semiconductors because you doubt AI capex, you cannot logically buy hyperscalers with the proceeds — they are the ones spending the capex. Those two positions express opposite views. The hyperscalers are exposed to whether the spend earns a return; the chipmakers are exposed to whether the spend happens at all.
On July 23 both fell. That's not a rotation, that's a repricing of the whole chain.
The mechanism
What a genuine rotation would look like: capital leaving the part of the chain with order risk and moving to the part with contracted revenue — that's power and electrical infrastructure, not cloud. What actually happened in July looks more like profit-taking in the most crowded, most-doubled corner of the market, followed by a broader risk-off day when the capex numbers landed.
There is a real distinction inside semis, though, and it matters more than the index move. Memory is priced on a commodity cycle. Logic and accelerators are priced on design wins and order books. A 18% index drawdown treats those as the same thing, and they aren't.
The risk
I could be wrong in a specific way: if hyperscalers cut capex, chips fall much further than cloud, and what looks like an irrational rotation now will look prescient. Order books are lagging indicators. By the time an order cut is announced, the stock has already moved.
Equally, a July drawdown after a first-half double is not evidence of anything. Semis do this. A violent 18% correction inside an uptrend is ordinary behaviour for the group, and reading a regime change into it is how people sell the bottom of a pullback.
My take
I'm not treating this as a rotation to follow. I'd rather use the drawdown to separate the two things the index conflates: businesses whose revenue is contracted against a buildout, and businesses whose revenue depends on a pricing cycle. The first group I'll hold through the noise. The second I'd only own with a plan to exit before the cycle turns — and after a first-half double, we're closer to the turn than the start.
Bottom line: selling chips to buy hyperscalers expresses two contradictory views at once. July was a repricing of the whole AI chain, not a rotation within it — and the useful distinction is contracted revenue versus cycle revenue, not chips versus cloud.
Not investment advice.