Same Chip Selloff, Wildly Different Pain—Here's Why TSMC Held Up

Same week, same "AI capex is peaking" panic, wildly different scoreboard. TSMC slid about 6.5% over the past week. Kioxia dropped nearly 14%. SK Hynix closed down over 14% in a single session, and on some trading days Samsung Electronics lost more than 13% in one shot. If you're holding Micron or TSM and wondering why one chip stock is bleeding twice as fast as the other, the answer isn't random — it's a structural difference in what these companies actually sell.

Photo by Andrey Matveev on Pexels

The Selloff Hit Everyone, But Not Equally

The trigger was broad: a roughly $1 trillion-plus rout across chip names as investors started pricing in the possibility that AI infrastructure spending is "peaking faster than expected." Reuters also pointed to forced unwinding of leveraged positions after SK Hynix's record profit still missed elevated Street expectations — a classic case of a beat that wasn't beat enough.

But TSMC just posted its fifth consecutive quarter of record profit, with Q2 revenue up 36% year-over-year to roughly $40.2 billion, beating consensus, and management raised full-year revenue guidance to over 40% growth. That's not a company signaling AI demand is cracking. So why did the stock still get dragged down with everyone else, just less violently?

Logic Versus Memory Is the Whole Story

TSMC is a foundry — it manufactures chips designed by Nvidia, Apple, AMD, and dozens of others, collecting margin on leading-edge process nodes regardless of which of those customers wins the AI race. Memory is a commodity business. DRAM and NAND are largely undifferentiated products sold on price, which means memory margins swing violently with supply and demand, and memory stocks get re-rated hard the moment sentiment shifts, leveraged bets or not.

That's exactly what I flagged in SK Hynix Posted a Record 76% Margin. Micron Fell 22% Anyway. — even genuinely great memory results don't protect the stock when the whole sector re-rates on macro fear. Logic doesn't escape selloffs entirely, but it has more customers, more use cases, and more pricing power to lean on.

China's CXMT Just Made the Memory Story Messier

Layer on top of the AI-capex jitters: CXMT, a Chinese DRAM maker, went public on Shanghai's STAR Market and rocketed over 465% on debut, hitting a market cap north of $85 billion. CXMT's global DRAM share jumped from roughly 3% to 8% in a year, and its capacity is on track to approach Micron's by year-end and hit 500,000 wafers per month by 2028 — about 17% of world DRAM supply.

That IPO alone knocked SanDisk down 12%, Micron down roughly 5%, and SK Hynix down 8% in a single session. It's a pure memory-sector threat — new commodity DRAM supply pressures pricing for exactly the kind of chips SK Hynix, Samsung, and Micron sell in bulk. TSMC has no direct read-through to a Chinese commodity DRAM entrant because it doesn't compete in that market at all.

Photo by PublicDomainPictures on Pixabay

Micron's One Real Moat: HBM

Here's the nuance that matters if you're deciding between TSM and MU right now. CXMT currently generates essentially no revenue from HBM (high-bandwidth memory), the specialized, high-margin chip that feeds AI accelerators. Micron holds roughly 21% of the HBM market. That segment is closer to TSMC's logic business in character — differentiated, technically hard to replicate, and not easily commoditized by a new low-cost entrant.

So the bear case against Micron right now is really a bear case against commodity DRAM and NAND, not against Micron's AI-memory franchise specifically. If you believe HBM demand holds up even as legacy DRAM gets squeezed by CXMT supply, the recent drawdown — which I covered in Micron's Down 31%. History Says the Drawdown Isn't Over Yet — starts to look like an overreaction on the blended-average business rather than a verdict on the part of Micron that actually matters most for its AI story.

What Would Actually Change This Read

This divergence only holds if AI infrastructure demand doesn't actually roll over. If hyperscaler capex genuinely slows — not just decelerates from "insane" to "very high," but actually contracts — TSMC's leading-edge logic orders get cut too, and the "TSMC is insulated" thesis breaks. Watch two things: TSMC's next quarterly guide (any walk-back from that 40%+ growth number is the tell), and whether CXMT's ramp starts showing up in DRAM spot pricing data over the next two quarters. Neither of those is visible yet — this is a read on current dynamics, not a forecast that's already been confirmed.

There's also a valuation risk worth naming: TSMC's stock already prices in near-peak margins and near-peak utilization, so even a modest demand wobble hits the stock harder than the fundamentals alone would suggest. "More resilient than memory" doesn't mean "cheap" or "risk-free."

The Takeaway

If you're deciding where to put new money in semis after this selloff, the TSM-versus-MU gap isn't noise — it's the market correctly pricing commodity-DRAM exposure differently from diversified logic exposure and specialized HBM exposure. Before adding to either position, pull up each company's next earnings call and specifically listen for capex guidance language and HBM allocation commentary — that's the data point that will tell you whether this divergence widens or closes.

This is not financial advice — always do your own research before making investment decisions.