The Chip Trade Changed Its Question — From 'How Much Capex?' to 'What's the Return?'

The most important sentence written about semiconductors this month is not about a chip. It is about a change in what investors are asking.

For two years, the reflex was simple: a hyperscaler announced bigger capital spending, and chip stocks went up. More capex meant more chips, and more chips meant more revenue for the companies selling them.

That reflex broke this summer. The SOX index fell more than 20% in July. Upstream suppliers — the memory and foundry names — came under visible pressure. And this week, when US chipmakers fell overnight, the damage travelled straight to Seoul and Tokyo, taking the KOSPI down 5.8% and the Nikkei about 3%.

The question has shifted from "how much are they spending?" to "what are they earning on it, and who is paying for it?"

The two forces doing the damage

Force one: the discount rate. Long-dated government yields have hit multi-decade highs across the US, Japan and Europe. Japan's 10-year touched roughly 2.945%, a three-decade high. A semiconductor company priced on years of future AI demand is the textbook definition of a long-duration asset. Raise the rate you discount those future cash flows at, and the present value falls — regardless of how good the business is.

That is why Samsung fell hard the same week it had record second-quarter profit and multi-year memory supply contracts in place. Nothing about the business deteriorated. The rate used to value the business changed.

Force two: the capex-return question. Investors are increasingly scrutinising whether AI infrastructure spending is generating commensurate revenue, and increasingly noticing that the marginal dollar funding it is borrowed rather than earned. We cover that shift in detail in our piece on how the AI trade moved from stocks to bonds.

Those two forces are the same force. Higher yields make debt-funded capex more expensive, which sharpens the return question, which lowers what investors will pay for the suppliers.

Why memory is the epicentre

Memory is the most cyclical corner of semiconductors. It has historically been a boom-bust business driven by supply cycles: prices spike, everyone builds capacity, prices collapse.

The AI era has been kind to memory because high-bandwidth memory is a genuine bottleneck for AI accelerators, and that has supported extraordinary pricing. But investors know the history, and cyclicality does not disappear because the end market is fashionable. When sentiment turns, memory falls furthest and fastest — which is exactly what happened this week.

The Korean example is instructive because index construction amplifies it. Samsung and SK Hynix together are roughly 47% of the KOSPI by market value. A memory selloff there is not a sector event, it is a national market event. We explain the mechanics — and which ETFs hold it — in our piece on Korea's two-stock index.

The signal in what didn't fall

Here is the most useful observation of the week, and it argues against panic.

On August 19, chipmakers fell — and the S&P 500 rose. Health care and cyclical stocks led. Moderna soared on trial results. Treasury yields dropped on the buyback announcement. Gold hit its best level since early June.

That is a rotation, not a market-wide risk event. Money is moving between sectors, not leaving. We wrote about this pattern earlier this summer in what "rotation" actually means, and it remains the single most useful concept for interpreting days like this one.

What a long-term investor should actually do

Check how much semiconductor exposure you have — across all your accounts. This is the practical action item. Semiconductors sit inside your S&P 500 fund, likely inside any Nasdaq-tracking fund, inside MSCI-based emerging-market funds via Korea and Taiwan, and inside any thematic AI or technology fund you own. Those add up quietly.

Do not treat a 20% index drawdown as a verdict. Sector indexes are volatile by construction. The SOX has had multiple drawdowns of this size during secular uptrends.

Do not add concentrated exposure because something fell. "It's down 20%, so it's cheap" is not analysis. The relevant question is whether the earnings that justified the old price are still coming.

Watch the capex guidance, not the stock prices. If hyperscalers cut capital spending plans, that is a real change in the demand picture for suppliers. So far, spending plans have broadly held — Amazon even raised guidance — which is why this looks like a valuation reset rather than a demand reset.

Frequently asked questions