美股

Has the August 2026 Memory Selloff Been Retraced? Same Three Sessions: SanDisk +14.8%, Samsung −8.6% — and 8.6 of Those 23 Points Are Just Seoul Being Closed

2026-08-04 · 11 min · AlphaGBM
闪迪SanDiskSNDK美光MicronMUSK海力士SK Hynix三星电子Samsung Electronics西部数据Western DigitalWDC希捷SeagateSTXHBMNANDDRAMHDD存储芯片memory chips半导体semiconductors板块轮动sector rotation跨市场比较cross-market comparison时区错配time-zone misalignment期权定位options positioningput call ratioSOXXSMHKRX市场结构market structure

Data as-of (important): U.S. figures are intraday prices at 10:30 ET on 2026-08-04 — not closing prices. MU / SNDK / WDC / SMH / QQQ come from Tiger real-time quotes; STX / SOXX / SPY come from a 15-minute-delayed source and are flagged as such. Korean figures are final 2026-08-04 KRX closing prices (that session has ended). The common baseline is each market's 2026-07-31 close, and every baseline close has been independently re-verified against historical K-line data (yfinance) and pykrx. The two markets' last prints are roughly 9 hours apart; Section 2 is devoted entirely to handling that gap rather than ignoring it.

Bottom Line

It was not retraced. It was reallocated. Over the same three sessions measured from the July 31 close, SanDisk is +14.79% and Samsung Electronics is −8.57% — a spread of 23.36 percentage points between companies that do the same thing: make memory chips.

But that 23.36pp cannot be taken at face value. 8.57pp of it exists only because Seoul has already closed and has not yet had a chance to price tonight's U.S. session. Net of the clock effect, the genuine same-information-set spread is 14.79pp. That decomposition is the core of this piece.

1. Three Sessions, One Industry, 23 Points Apart

Baseline = each market's 2026-07-31 close.

Instrument Leg 7/31→8/03 close 8/03→now 7/31→now Basis
SanDisk (SNDK) NAND-weighted +6.03% +8.26% +14.79% US intraday
Micron (MU) DRAM / HBM +0.79% +6.75% +7.59% US intraday
Western Digital (WDC) HDD −3.23% +3.46% +0.12% US intraday
Seagate (STX) HDD −2.93% +2.63% −0.37% US intraday · delayed
SK Hynix (000660) Highest HBM purity −8.79% +0.64% −8.21% KRX final close
Samsung Elec (005930) HBM + diversified −8.76% +0.21% −8.57% KRX final close

Raw prices for replication: 7/31 closes — SNDK 1,214.83 / MU 823.03 / WDC 544.84 / STX 856.13; Hynix KRW 1.718M, Samsung KRW 262.5K. 8/04 KRX closes — Hynix KRW 1.577M, Samsung KRW 240.0K. Current U.S. — SNDK 1,394.48 / MU 885.48 / WDC 545.47 / STX 852.95.

2. Subtract the Clock First: 8.57 of Those 23.36 Points Are an Artifact

This is the single most common error in cross-market comparison. Seoul closed at 15:30 KST on August 4, which is 01:30 ET. The U.S. readings in this article are from 10:30 ET. In other words, Seoul's last trade printed a full nine hours before the U.S. last trade, with tonight's entire U.S. session in between.

Subtracting one from the other holds Korean stocks responsible for a tape they never had the chance to trade against. The clean decomposition:

Comparison basis SanDisk Samsung Spread
A. Both through the 8/03 close (same information set) +6.03% −8.76% 14.79pp
B. Including tonight's U.S. intraday (different information sets) +14.79% −8.57% 23.36pp
B − A = increment Seoul has not yet priced 8.57pp

Conclusion: 8.57pp of the 23.36pp — roughly 37% — is pure time-zone artifact, not valuation disagreement. The real disagreement is line A's 14.79pp.

That increment has not vanished; it is queued for the next Seoul session. How it gets resolved is something this article does not forecast.

3. Net of the Clock, the Remaining 14.79pp Is Structural

The decisive point: the ordering held in both the selloff and the rebound. It never inverted.

On August 3, when everything fell, the ranking was SNDK > MU > STX > WDC (least to most damaged, with the Korean pair last). On August 4, when everything rebounded, the ranking was again SNDK > MU > WDC > STX. An ordering that repeats in two opposite directions is not noise — it is a pricing mechanism.

Sorted by HBM exposure from low to high, three-session cumulative returns are nearly monotonically decreasing:

Reading: the market is not repricing "memory demand." It is repricing "HBM content per system." Had aggregate memory demand been the thing sold, the two HDD names would not be sitting flat, and SanDisk certainly would not be posting a double-digit gain. Memory is no longer one block — it is three legs priced separately, and that dispersion is the most tradeable factual feature of this episode.

An honest note on attribution: the market has linked this drawdown to a rumor that NVIDIA's Rubin Ultra will cut HBM specification (12Hi→8Hi), relayed by Jukan citing Eugene Investment Securities. We could not independently verify that rumor and do not rely on it. A directly contradicting claim circulated in the same window: DigiTimes reported 2027 DRAM/HBM capacity is already sold out, and Jukan himself framed the issue as supply-side rather than demand-side. Two mutually exclusive rumors circulating at once is itself one source of the volatility. The return ordering above is a price fact and does not depend on either rumor being true.

4. Seoul Traced a V Today — and Our Own Morning Call Was Falsified the Same Day

For transparency, on the record: our August 4 morning read was wrong, and was falsified within the session.

Roughly 8 percentage points of intraday range that closed flat: neither side won. That is not the shape of distribution. Inferring flow intent from intraday readings is unreliable in a tape with this much range; only final prints constitute evidence. This was our own tuition payment this cycle, recorded here so readers can discount our judgment accordingly.

5. Options Positioning Runs Exactly Opposite to the Price Ranking

On August 4 pre-market options data, the best performer is the one nobody believes:

Instrument Put/Call 30-day average Reading
SanDisk (SNDK) 1.74 1.53 Strongest of the three legs — yet buying protection
Micron (MU) 0.294 1.123 (z ≈ −1.7) Trails SanDisk by 7.2pp over three sessions — yet positioned extremely bullish

"Nobody believes the winner; everybody is crowded into the laggard" is the signature of this positioning. Micron is simultaneously "someone's high-conviction bet" and "a bet that is too full" — the most crowded name in the group. Crowding does not indicate direction, but it does mean that if the thesis is falsified, price will adjust faster than fundamentals change.

6. Semis vs. Market: This Is Rotation, Not a Broad Rally

Same three sessions, same 7/31 baseline:

Comparison Semis Broad market Excess
SOXX vs SPY +5.61% +2.32% +3.29pp
SMH vs QQQ +5.21% +4.07% +1.14pp
SMH vs QQQ (today's session only) +4.26% +2.28% +1.99pp

Over the same window, mega-cap technology was flat to lower. The semiconductor excess return is not being pushed by new money; it is existing money moving between sectors. That makes its durability a function of how much can still be pulled out of other sectors, not a function of aggregate liquidity.

7. How to Replicate This Article

Every number here is verifiable from public data in minutes:

  1. U.S. baseline and 8/03 closes: pull MU / SNDK / WDC / STX closes for 2026-07-31 and 2026-08-03 from any quote source and check them line by line against the raw prices in Section 1.
  2. Korean data: pykrx, via get_market_ohlcv('20260731','20260804','000660') and '005930', closing-price column.
  3. Cross-validation: computed independently off the 7/31 baseline, Samsung's 7/31→8/04 return here is −8.57%, matching the −8.57% we recorded independently on the evening of August 4; the 8/03 percentage moves for all six names (+6.03 / +0.79 / −3.23 / −2.93 / −8.79 / −8.76) likewise reconcile line by line with same-day records. Two independent paths producing the same set of numbers is the primary basis for the reliability of this data.

Methodological Limits (please read)

想要更深入的期权数据与实盘分析?了解 AlphaGBM →