ROGUEQUANT_

$ cd ..

2026.07.20 // MARKETS // 3 MIN

The July 2026 Chip Selloff: First Real Stress Test of the AI Trade

The SOX fell 24% in three weeks. Intel lost a fifth of its value, Korea hit circuit breakers, and Samsung dropped 7% on a +1,800% earnings print. What the semiconductor correction does—and doesn't—tell you about the AI trade.

ALESSIO ROCCHI ·

Two weeks ago I wrote that the AI trade was crowded and priced for perfection, and that you didn't need to predict the pin—just avoid being the most crowded person in the room when it found one. The room just got its first evacuation drill.

The Philadelphia Semiconductor Index dropped roughly 24% from its late-June all-time high in three weeks, a technical bear market. This piece is about what actually broke—and what didn't.

The Tape

FIG. 01 // THE TAPE

Drawdowns from recent highs, July 2026 chip selloff

SOX INDEX: −24%SOX INDEX−24%INTEL: −21%INTEL−21%MICRON: −17%MICRON−17%SK HYNIX: −12%SK HYNIX−12%SAMSUNG: −9%SAMSUNG−9%
The Philadelphia Semiconductor Index entered a technical bear market three weeks after its all-time high. The KOSPI fell ~10% intraday at the worst point, triggering circuit breakers.

DATA: REUTERS · CNBC · BENZINGA · JUL 2026

The damage is broad and international: Intel −21%, Micron −17%, and a session where Samsung and SK Hynix fell 9-12% and the KOSPI dropped nearly 10% intraday, triggering circuit breakers. European semis followed—ASML, STMicro, Infineon all down 4-5% in sympathy.

Note what's absent from the tape: credit stress. Investment-grade spreads on the hyperscalers barely moved. This was an equity-positioning event, not (yet) a financing event—that distinction is the whole ballgame, and it's the first thing to check in any AI drawdown, as argued in the capex piece.

The Paradox That Explains Everything

FIG. 02 // PRICED FOR PERFECTION

When record earnings meet a crowded trade

Samsung Q2 operating profit, year-over-year

+1,800%

89.4 trillion won — a record print

Samsung stock reaction to that print

−7%

beating estimates was not enough

SOX: first-half run → three-week drawdown

+65% → −24%

the exit was narrower than the entrance

A stock that falls 7% on a +1,800% earnings print is not trading on earnings. It is trading on positioning — the defining feature of this correction.

DATA: SAMSUNG IR · CNBC · JUL 2026

Samsung reported a record ~89.4 trillion won quarterly operating profit, up more than 1,800% year-over-year—and the stock fell 7%. Read that twice. A company beat a high bar by a mile and was sold anyway.

A stock that falls on record earnings is not trading on earnings; it's trading on positioning. After a 65% first-half run, everyone who wanted to be long semis was already long. When marginal buyers are exhausted, good news has no one left to convince—and the exit is always narrower than the entrance.

Three Ordinary Catalysts

FIG. 03 // THE CHAIN

Three ordinary catalysts, one crowded exit

META COMPUTEsurplus AI capacity for saleINTEL 18Aprofitable yields slip to 2027SK HYNIXHBM4 expansion delayedPRICED FORPERFECTION−24%IN 3 WEEKS
None of the three catalysts breaks the AI-infrastructure thesis on its own. Together they hit a sector up 65% in six months — and positioning, not fundamentals, did the rest.

SOURCES: AXIOS · FORBES · CNBC · JUL 2026

What actually triggered it? Three announcements, each unremarkable in isolation:

  1. Meta Compute: Meta plans to sell surplus AI training and inference capacity to enterprises. The first hyperscaler admitting to surplus punctures the "insatiable demand" premise—though it can equally be read as a compute spot market being born.

  2. Intel 18A: profitable yields on Intel's flagship node slip to late 2026 or 2027—a company-specific execution miss.

  3. SK Hynix: HBM4 expansion delayed in favor of DDR5—read by the market as moderating AI memory demand; equally readable as margin discipline.

None of these breaks the AI-infrastructure thesis. All three hit simultaneously, into record concentration, with rate futures now pricing a Fed hike as soon as October. Mechanical repricing did the rest.

What Would Actually Confirm the Bear Case

Discipline requires stating falsifiers in advance. The selloff becomes fundamental if you see:

  • Capex guidance cuts from a second hyperscaler (Meta selling surplus is one data point; two is a trend)
  • The megawatts not showing up: interconnection-queue withdrawals and cancelled PPAs—the honest demand signal from the power piece
  • Credit spreads widening on data-center-heavy issuers while equities bounce—debt markets sniffing out what equity denies
  • Circular-deal write-downs: any impairment inside the Nvidia-OpenAI-Oracle loop propagates to all of it at once

Until those trigger, a 24% drawdown after a 65% run is what crowded trades do on schedule: Yardeni's "another 12% possible" and "valuation, not fundamentals" can both be true.

The honest assessment: this correction validated the mechanics of the bubble thesis—concentration, crowding, reflexivity—without yet validating its substance. The dispersion trades flagged two weeks ago paid; the falsifiers above are what would turn a positioning washout into a regime change. Watch the queues and the spreads, not the headlines.

Did your risk models see the KOSPI circuit breakers coming, or did correlation assumptions break first? Backtests tell one story; the tape tells another.