Why Are Memory Stocks Up ~5% While the S&P 500 Is Down on September 4, 2026? Three Reproducible Tests for Sector vs. Market Moves

Data basis (read this first)

• Korean figures: settled closing values from the Korea Exchange (KRX) session of 2026-09-04.
• U.S. figures for 09-03: settled close of 2026-09-03.
• U.S. figures for 09-04: intraday readings taken at 10:32 ET on 2026-09-04 — that session had not closed. Intraday data moves. Every statement resting on the 09-04 U.S. session is labeled "intraday," and explicit post-close falsification conditions are given at the end.
• Hong Kong 09-04 had not entered our readings layer (CCASS settles T+1; the options snapshot lands at 22:00 Beijing). No Hong Kong 09-04 figure is cited here.

Every ratio below can be recomputed from the raw share counts, volumes and percentages given in the text.

The short answer

Because this is a sector move, not a market move. At the same instant the broad index and the memory group carry opposite signs — the S&P 500 ETF is down 0.30% while the four memory names are up an average of 5.29%, a spread of 5.59 percentage points. When the broad index falls and one group rises 5%, whatever explains it must be specific to that group. A market-wide variable such as "sentiment improved" cannot do the work.

Further: the move has a clear internal ordering, and the ordering itself carries information — the NAND/HDD names lead the DRAM name by 1.49 percentage points, which points at enterprise storage rather than HBM.

Test 1: Broad index and sector carry opposite signs → rules out a market move

Intraday readings, 2026-09-04, 10:32 ET (session not closed)

Three recomputable spreads

• Memory four average − S&P ETF: (7.45 + 5.04 + 4.50 + 4.17) ÷ 4 = +5.29%, minus (−0.30%) = 5.59 percentage points
• Memory four average − Nasdaq-100 ETF: 5.29% − 0.18% = 5.11 percentage points
• Semiconductor ETF − S&P ETF: 2.57% − (−0.30%) = 2.87 percentage points

Why this test works

If the driver were a market-wide variable — rates, macro data, risk appetite — the broad indices would have to move in the same direction. Right now the S&P is negative, the Nasdaq-100 and Russell are near zero, and all four memory names are up more than 4%. A market-wide variable cannot explain "only one cluster is up 5%."

Note Broadcom at −0.11%: it is a semiconductor, but not a memory name. The semiconductor ETF's +2.57% is carried by the memory and logic/AI constituents, not by a broad industry rally. Dispersion exists inside the sector too.

⛔ This article does not attribute the session to news. We did not verify a first-hand event source that would explain four names gapping together. "No catalyst found" is not the same as "no catalyst exists" — but it is also not a licence to invent one. What follows uses only recomputable prices, volumes and ownership structure.

Test 2: Ordering inside the sector → NAND/HDD leads DRAM

Split the four by business centre of gravity:

Why the ordering matters

For the past two years the default narrative for memory rallies has been anchored on HBM and AI training memory. If that were the leg moving today, Micron should lead — it has the most direct HBM exposure of the four. The opposite holds: Micron is the weakest of the four, and the leaders are NAND and mechanical hard drives.

Seagate and Western Digital have no HBM exposure at all, yet both are up more than 4.5%. That compresses the explanatory space toward nearline storage, enterprise capacity and the data-centre storage tier — not AI training memory.

This test is falsifiable: if the ordering inverts over the next few sessions, with Micron leading and the HDD names lagging, the "enterprise storage leg" reading fails and the HBM narrative should be restored.

Test 3: Asia moved first, and the slow variable followed (the strongest evidence here)

This is the only chain built entirely on settled, closed-session data, so it carries the most weight.

The sequence

Asia moved first; the U.S. session followed roughly eight hours later with greater magnitude. That is a timestamped, verifiable cross-market sequence.

The slow variable did follow — SK Hynix

Price can be pushed by intraday round-tripping. Foreign ownership ratio cannot — buying and selling within the day leaves no trace in it. That makes ownership the key slow variable here.

• 09-04 foreign net buying of 308,833 shares, equal to 10.92% of session volume
• Domestic institutions +96,776 shares, moving in the same direction as foreigners
• Volume 2.828M shares, the highest of the three sessions → the advance came on rising volume
• Foreign plus institutional combined: 09-02 −KRW 2.60tn (about −$1.92bn) → 09-03 −KRW 0.86tn → 09-04 +KRW 1.27tn (about +$0.94bn)

USD figures use the live 2026-09-04 exchange rate, not a fixed constant.

Same direction, completely different buyers — Samsung

The conclusions must be drawn separately:

• SK Hynix: global allocation money added back a little. The evidence is that the slow variable — ownership — moved with it.
• Samsung: the bid came from domestic institutions alone. They bought 2.49M shares — eight times the size of the foreign buying in Hynix — yet foreign ownership barely moved, because foreigners were net sellers in the same session.

So at this point the honest statement is that the buyer structure has diverged. It is not "foreign investors are back." The Hynix pattern (foreign buying plus a rising ownership ratio) is typically the more persistent of the two, but its absolute size is a fraction of Samsung's.

Relative strength: one session of outperformance is not a trend

Hynix versus Samsung: −0.71pp (09-02) → −0.85pp (09-03) → +1.00pp (09-04, first outperformance). But cumulatively across the three sessions Hynix is −2.71% against Samsung's −2.10%, leaving Hynix still behind by 0.61 percentage points. The outperformance happened in exactly one session.

The catalyst search came back empty — and empty means genuinely absent

DART (Korea's Financial Supervisory Service electronic disclosure system), window 08-28 to 09-04: two filings for Hynix, four for Samsung, zero in the material category, query status normal. That is "there genuinely were no material disclosures," not "we could not check." ⛔ Accordingly, no news-based story may be attached to the Korean leg.

One piece of contrary evidence (we are not hiding it)

Publishing the unfavourable evidence alongside the favourable is what makes a judgment checkable.

The 09-03 block-trade flows point the opposite way from 09-04 price action.

In the 09-03 session, block flow was buying Micron and selling SanDisk, with SanDisk the second-largest outflow on the entire tape. The next session, intraday, SanDisk is the strongest of the four, outpacing Micron by 3.28 percentage points.

How to read this:

1. One-day block net did not predict next-day direction. At least not this time. Any mechanical use of "block inflow equals bullish" has a counterexample here.
2. But keep the definitions straight: the criterion we wrote down on 09-03 was "Micron block net minus SanDisk block net = +$158.3M," and it tests whether the flow itself persists, measured at the 09-04 U.S. close. It is not a price criterion, its measurement point has not arrived, and this article claims neither confirmation nor refutation of it.
3. A more neutral reading: on 09-03 prices barely moved (Micron +0.22%, SanDisk +0.10%) while ownership turned over heavily — turnover is easiest to miss precisely when price does not move. The next-day gap may be the delayed pricing of that turnover. This is a hypothesis, not a conclusion.

What would falsify the above (graded, covering the full range)

On "sector move, not market move"

Criterion: memory four average minus S&P ETF change (5.59pp intraday today), measured at the 09-04 U.S. close:

On "NAND/HDD leads DRAM"

Criterion: NAND/HDD trio average minus Micron (+1.49pp intraday today):

On "foreign investors rebuilding SK Hynix"

Baseline: foreign +308,833 shares / ownership 50.49% / volume 2.828M shares. Measured at the next Korean session:

On Samsung's "domestic-only bid"

Baseline: institutional +2,489,812 shares, foreign −115,377 shares.

• Institutions buy ≥ 1M again and foreigners turn buyers → the two legs converge; upgrade to pending verification
• Institutions buy while foreigners still sell → "domestic-only bid" stands, and this state may not be used to support either side
• Institutions turn net sellers ≥ 1M → settlement-driven buying; void

Definitions and limits (please read in full)

1. The U.S. 09-04 session had not closed. Everything said about it is provisional and the close may change the conclusion. The falsification table above exists for exactly this reason.
2. An implied-volatility percentile is not a "12-month low." Our options sample has been self-collected since 2026-03-20 and spans only 85–114 trading days, not 252. The current S&P ETF ATM IV percentile of 0.9 and Nasdaq-100 ETF 0.0 (n=114) may only be described as a floor within the self-collected sample.
3. No high/low or valuation call is made here — the fourth consecutive session without one. The valuation source is as-of 08-29, six days behind (it refreshes every eight days); memory names are cyclicals, where the P/E is highest at the trough of earnings and the percentile direction inverts entirely; and we currently have no price-to-book percentile tool. No call is made, and drawdown depth is not substituted for valuation.
4. The four companies have different fiscal year-ends, which makes a static cross-company P/E comparison wrong by construction. No such comparison appears here.
5. Claims without a verified first-hand source are excluded from the evidence chain. For example, the circulating claim that "Micron is doubling high-bandwidth memory capacity" is a third-party paraphrase for which we have not located a primary source, so it is not used — and since it would cut against the two Korean names directionally, it certainly cannot be used to explain today's advance.
6. Hong Kong 09-04 had not entered the readings layer (CCASS is T+1, the options snapshot lands at 22:00 Beijing). This is a publication cadence, not a data fault, and nothing from it is cited.
7. The DRAM spot price field is permanently empty for us (TrendForce / DRAMeXchange require a paid subscription). No spot quote is used anywhere in this article.
8. No news attribution is offered. We did not verify a first-hand event source explaining the simultaneous move in the four U.S. names on 09-04, so no cause is asserted — only a recomputable structural description.

Frequently asked questions

Q: Memory stocks are up — does that mean semiconductors are up as a sector?

No. Broadcom is −0.11% at the same moment, and the semiconductor ETF's +2.57% is carried by the memory and logic/AI constituents. A rising sector ETF does not mean its constituents rose broadly. You have to decompose it.

Q: The S&P is down and memory is up — is money rotating out of the index into memory?

This article does not make that claim. "Rotation" requires flow evidence, not merely opposite price signs. The only settled flow evidence available is the Korean leg (Hynix foreign net buying of 308,833 shares, ownership 50.44%→50.49%); U.S. flow data for 09-04 does not exist until the close. Opposite price signs prove the driver is sector-specific; they do not prove money moved from A to B.

Q: Why is Micron up the least?

That is exactly the phenomenon Test 2 addresses. Micron has the most direct HBM exposure of the four, so if the AI-memory leg were the one moving, it should lead. The actual ordering has NAND/HDD ahead by 1.49 percentage points, pointing at enterprise storage rather than HBM. This is an observation, not a cause — we did not verify a catalyst.

Q: Foreign investors turned buyers in Korea — are foreign investors back?

No. Only the Hynix leg shows foreign buying with a rising ownership ratio (50.44%→50.49%). In Samsung, domestic institutions bought 2.49M shares while foreigners were net sellers of 115,377 shares in the same session and ownership barely moved. One session is not a trend, and the Hynix amount is a fraction of Samsung's. The honest statement is that buyer structure has diverged.

Q: Block flow was selling SanDisk on 09-03 — why was SanDisk the strongest the next day?

That is the contrary evidence published above. At least in this instance, one-day block net did not predict next-day price direction. One possible — and unverified — explanation is that prices barely moved on 09-03 while ownership turned over heavily, and the next-day gap is the delayed pricing of that turnover. A hypothesis, not a conclusion.

This is public-layer market-structure analysis. It contains only verifiable public data, methods and criteria. It contains no holdings, position sizes or trading actions, and is not investment advice. Every ratio can be independently recomputed from the raw share counts, volumes and percentages provided.

Generated 2026-09-04 22:32 Beijing time / 10:32 ET. The U.S. session was open at the time of writing.