Memory Stocks Trailed the Semiconductor ETF by 2.80 Points on Sept 8, 2026 — Mispricing or Regime Change? A Pre-Registered Test Fired in the First Hour of Sept 9

Data basis (read this first)

• U.S. figures: ⚠️ U.S. markets were still open at the time of writing. Same-day U.S. numbers are live intraday prices for 2026-09-09, as-of 10:48 ET (approximately one hour into the 09:30–16:00 ET session, with roughly five and a half hours remaining). Intraday prices will change. The comparison basis is the 2026-09-08 settled close.
• All figures labeled "09-08" are settled closing values and do not move with the live session.
• Korean figures: settled closing values from the Korea Exchange session of 2026-09-09.
• Cross-market: ⛔ This article computes no U.S.–Korea price spread. Korea has closed while the U.S. session is live; the as-of basis differs, so any ratio between them would be invalid. We state sequence only, never spread and never causation.
• Hong Kong: not cited (the 09-09 session did not enter our reading layer).

1. The conclusion first

At the 2026-09-08 U.S. close, memory names were left behind by the semiconductor index. The semiconductor ETF (SMH) closed +1.19%, while in the same session Micron (MU) closed -1.61% and SanDisk (SNDK) -0.12% — trailing SMH by 2.80 and 1.31 percentage points respectively.

Two mutually exclusive explanations were on the table:

• Explanation A (mispricing): 09-08 was single-session noise, and memory remains part of this semiconductor move.
• Explanation B (regime change): money did not leave semiconductors, but it did leave memory. The leadership rotated to compute design and manufacturing (AMD / Broadcom / TSMC), and memory is the slot being routed around.

These two explanations could not be separated using 09-08 data, because they explain 09-08 equally well. So the test was pre-registered before the next session opened: observe how Micron and SanDisk perform relative to SMH on 09-09. Both outperform = A. Both underperform = B. One of each = flash and DRAM are still diverging internally, and the two must not be judged together.

As of 10:48 ET (roughly one hour into the session), Explanation A is the branch that fired — and not marginally.

⚠️ These are intraday readings, not settled values. The pre-registered test requires settlement at the close. What follows is a mid-session score, and the final verdict must wait for the 16:00 ET close. Section 6 defines the test precisely and states the conditions under which it flips.

2. The key point: this is not "the market went up and memory followed"

Separating "memory is moving on its own" from "a rising tide" requires only the broad benchmarks in the same session:

• Nasdaq 100: -0.04% — essentially flat
• Semiconductor ETF: +0.28% — marginally higher
• Micron +3.69%, SanDisk +3.09% — substantially ahead

The broad index is flat, the semiconductor index is barely up, and the two memory names are each generating two to three percentage points of excess return. That means today is not memory being lifted by a general recovery in risk appetite. It is capital being reallocated inside semiconductors — in the opposite direction from 09-08.

Within the same session, 09-08's leaders are confirming this in reverse:

Broadcom, the biggest gainer on 09-08, is down today. Memory, which fell on 09-08, is up. Seats were swapped between the two sessions, while the broad index barely moved across both (Nasdaq 100 two-session cumulative: -0.119%).

3. Across both sessions, the 09-08 gap has already closed

A single session is easy to misread. Compounding the two is cleaner (09-08 settled × 09-09 intraday):

• Micron's two-session cumulative is +2.020% against SMH's +1.470%, a gap of +0.550 pp — the 2.80 points it gave up on 09-08 were erased within the first hour.
• SanDisk's two-session cumulative is +2.963%, now +1.492 pp ahead of SMH.

Put differently: if at the 09-08 close all you saw was "memory is lagging," then one hour into 09-09, that lag no longer exists in the data. That is the definition of the mispricing branch.

4. The irony on the options side: the only name that was insured is today's biggest gainer

The 09-08 settled options positioning is the single most notable data set here:

In the 09-08 session, the market chased Broadcom, AMD and TSMC while buying almost no protection on that rally — and simultaneously bought insurance on Micron as it fell. A put/call ratio above 1 means Micron's put positioning exceeded its call positioning that day, which was unique among the nine names tracked.

One hour into 09-09, the insured name (Micron) is +3.69% and the uninsured one (Broadcom) is -1.52%.

This does not establish any general "options are a contrarian indicator" claim. A single day proves no rule, and puts may be one leg of a covered or spread position rather than an outright bearish bet. It is worth recording for a different reason: on 09-08, the money chasing the rally and the money buying protection priced the same sector in opposite directions, and the 09-09 open sided with the protection buyers.

Two further settled structural facts belong here:

• Protective positioning sits on the index leg, not the single-stock leg. 09-08 settled put/call ratios: SMH 3.137, Nasdaq 100 2.294, S&P 500 1.44, Hong Kong Tracker Fund 11.961. On the single-stock leg, every name except Micron (1.105) is below 1. The index has a cushion; AMD and Broadcom are naked.
• Implied volatility percentiles are broadly in a low band. Eight of the nine names we track sit below the 25th percentile on implied volatility (Broadcom 4.4, Micron 6.1, SanDisk 7.0). ⚠️ Methodological limit: this percentile is computed over a self-collected sample of 88–117 trading days, not 252 trading days. It therefore must not be read as a "one-year low" — only as "low within our collection window."

5. The Korean leg: SK Hynix rose first, but the attribution does not hold

In sequence, the Korean 09-09 close occurred before the U.S. 09-09 open:

• SK Hynix closed at KRW 1,856,000, +3.51%
• Samsung Electronics closed at KRW 269,500, 0.00%
• Hanmi Semiconductor closed at KRW 250,500, +2.45%

⛔ This must be stated plainly: the attribution for the SK Hynix gain does not hold, and we record it as unattributable. The day's flow structure supports no clean explanation:

• Foreign investors were net buyers for a third consecutive session, but the size is collapsing: 772,000 → 97,500 → 41,700 shares across the three sessions (contractions of 87.4% and 57.2%).
• Foreign plus institutional flow accounted for only 3.31% of the day's turnover, and the foreign ownership ratio moved only from 50.66% to 50.67%.
• In other words, +3.51% was produced with minimal foreign participation, so pricing power did not sit on the foreign side — and our reading layer does not decompose the remaining 96.7% of turnover.

A same-day disclosure must also be reported accurately, and explicitly treated as not constituting an explanation:

On 2026-09-09, SK Hynix filed a reply to an exchange inquiry regarding a rumor or media report (disclosure ID 20260909800559). The document specifies a re-reply deadline of 2026-12-08.

⛔ The content of that reply is not in our reading layer and we were unable to verify it independently. This article therefore reports only that such a filing exists, that its nature is undetermined, and that the deadline is 12-08. We do not speculate on its content, and we do not treat it as the cause of the day's +3.51%. Attaching a filing of unknown content to a price move is textbook post-hoc narrative.

In the same session, Korean non-semiconductor names rebounded sharply (Samsung SDI +8.30%, LG Energy Solution +6.46%) — but these were among the largest decliners in the prior session, so the move carries an inherent oversold-bounce component and cannot be read directly as sector rotation.

Accordingly, this article does not claim a "Korea rises, then the U.S. follows" transmission chain. We state two settled facts and their ordering, and leave causation blank.

6. How the test was pre-registered, and what would flip it

This section is the only part of this article that adds anything beyond restating price action, so it is spelled out in detail.

A test must be defined before the outcome is known, or it is merely post-hoc narrative. The version written down after the 09-08 close was mutually exclusive, falsifiable, and exhaustive:

The three branches exhaust the outcome space, leaving no gap where "none apply." That design point was learned the hard way: a separate test written the same evening, covering Samsung Electronics, only specified branches for "continued net buying, bucketed by share of turnover" and failed to cover a sign flip. When the actual result flipped to net selling, neither branch applied and the entire test was void. That was a flaw in test design, not a data problem.

Current status and falsification conditions:

• Now (10:48 ET, roughly one hour in): the first branch has fired, and both names' excess return over SMH exceeds 2.5 percentage points — not a marginal trigger.
• ⚠️ Not yet settled: the test requires settlement at the 16:00 ET close. If either Micron's or SanDisk's excess return over SMH converges to negative by the close, the verdict falls into the third branch (one of each, do not judge together). If both turn negative, the second branch holds and the conclusion in Section 1 should be withdrawn in full.
• How many sessions would it take to matter: a single outperforming session can only reject "memory is being systematically routed around." It cannot establish "memory is leading again." That would require sustained excess return across multiple sessions, formed on expanding rather than contracting volume.

7. Reviewable raw readings

2026-09-08 U.S. close (settled, reviewable)

2026-09-09 U.S. intraday (as-of 10:48 ET, unsettled, will change)

2026-09-09 Korea Exchange close (settled)

8. What this article explicitly does not claim

Listed individually to prevent selective quotation:

1. No claim that the memory cycle has bottomed or reversed. This article addresses one narrow question — whether the 09-08 decline was a mispricing. Two sessions cannot speak to a cycle.
2. No claim of a Korea-to-U.S. transmission chain. Sequence is stated; causation is left blank.
3. No attribution for SK Hynix's +3.51%. Recorded as unattributable.
4. No speculation on the content of the exchange inquiry reply.
5. No U.S.–Korea price spread is computed. The as-of bases differ.
6. Implied volatility percentiles are not read as "one-year lows." The sample window is 88–117 trading days.
7. No positions, position sizing, or trading recommendations. This is a public-layer record of data and a pre-registered test.

Sources: live U.S. quotes via the Tiger Brokers API (as-of 2026-09-09 22:48 CST); 09-08 U.S. settled closes and options positioning, and the 09-09 Korea Exchange closes and investor-type ownership ratios, from that day's retained reading-layer files. All percentages are computed from raw prices with no processing beyond rounding.

as-of timestamp: 2026-09-09 22:48 CST / 10:48 ET. Same-day U.S. figures will continue to change until 16:00 ET.