Why Chip Stocks Are Down Today (It’s Not the AI Trade)

Everyone Calm Down: The Chip Selloff Is a Rates Story, Not an AI Story

Our screens are a sea of red. AMD off almost 4%. Lumentum and Coherent down 5%. The semis index gave back 5% in a single session and the small-cap photonics names are bleeding twice as hard as the ETFs. Look at nothing but the tape and you’d swear somebody canceled artificial intelligence.

Nobody did. Go read the actual news flow. Not one of these companies reported anything. No guidance cuts, no canceled orders, no hyperscaler pulling capex. The selling has nothing to do with chips and everything to do with the bond market.

What Actually Hit the Tape

The 10-year is sitting near its highest yield since early 2025. Inflation prints won’t roll over, Washington keeps issuing paper like it’s free, and now oil is parked around $85 while the US and Iran stare each other down over the Strait of Hormuz. Energy feeds inflation, inflation feeds yields, and yields are kryptonite for growth stocks. The selling started in Asia overnight — Chinese chip names got hit on the same yield surge — and washed right back into our open.

That’s a discount-rate problem, not a semiconductor problem. When yields back up, the market marks down future earnings, and the further out those earnings live, the harder the markdown. Math, not a verdict.

The Tape Confirms It

Here’s the tell, and it’s hiding in plain sight: look at the ordering of the damage. Nvidia closed down less than 1%. Taiwan Semi, less than 1%. The two companies with actual AI revenue today barely got scratched. Now walk out the duration curve — Micron off 1.6%, SOXX down 2.3%, AMD down 3.7% — and keep walking until you hit the pre-revenue photonics names taking 4% to 6% body shots.

That’s exactly what a rate shock looks like. A demand problem would run the other direction — Nvidia would be the epicenter, because Nvidia is where the AI revenue lives. Instead the hardest-hit names are the ones priced on 2028 and beyond. The market isn’t saying the future got worse. It’s saying the future got more expensive to wait for.

There’s a mechanical layer under this too. When a crowded sector reverses, levered funds get carried out on stretchers whether they like their positions or not. Jim Cramer flagged it during July’s chip rout, calling the sellers “monstrous, motivated and often margined.” Forced selling doesn’t read 10-Ks. It answers margin calls. Ugly to sit through, and it tells you nothing about where these businesses are in three years.

What Didn’t Change This Week

While the tape melted, the fundamental news in photonics ran the other way. A new China Insights Consultancy report built on LightCounting data projects the AI data center optical interconnect market growing from $13.7 billion in 2024 to $144 billion by 2030 — a tenfold expansion in six years, with silicon photonics expected to take roughly two-thirds of it. That report dropped this week. During the selloff.

Zoom out and the setup keeps getting better. The consensus coming out of OFC 2026 was that essentially every high-bandwidth data center interconnect goes optical within five years — Nvidia has co-packaged optics on the roadmap for scale-up around 2028, and Meta showed reliability data favoring CPO over pluggables. Meanwhile the Street has started framing optical interconnect capacity as the next bottleneck in the AI supply chain. Bottleneck is the magic word. Bottlenecks are where pricing power lives.

Company level, same story. POET Technologies just printed its sixth straight quarter of sequential revenue growth, up 112% year over year, with $796 million in cash after the May financing, and production units are shipping for qualification into year-end — the execution case we laid out before the print got answered line by line. A company sitting on that much cash doesn’t need to raise into weakness, which kills the dilution objection retail reflexively reaches for. None of that got revised this week. The tape did.

Our Take

We’ve traded through enough of these to know the choreography. Rates spike, duration gets puked, the babies go out with the bathwater, and six months later everyone swears they were buying the dip the whole time. The hyperscalers aren’t pausing data center construction because the 10-year backed up. The inference workloads eating all this bandwidth don’t care that oil touched $85.

So ask the only question that matters: did anything happen this week that reduces how much data AI systems will need to move in 2028? No. Which means this selloff repriced the waiting, not the destination. Volatility is the toll you pay to own the buildout. Some days the toll booth charges more. The road still goes where it goes.

None of that means the drawdown is finished. Rates can keep climbing, Hormuz can get uglier, and forced sellers stay forced longer than seems rational. But there’s a difference between a thesis breaking and a thesis going on sale. This is the second one.