Is the NAND Rally Still Running? The Storage-Purity Ordering Fell From +0.92 to Roughly Zero, and Why Samsung Overtook SK Hynix

Data basis (read this first)

• Two settled reference points: U.S. equities at the 2026-09-18 close (the last settled U.S. session at the time of writing), and Korea at the 2026-09-21 close. The core conclusions rest on these two.
• The 09-21 U.S. session was trading while this was written. Every figure labeled "09-21 intraday" was captured at 2026-09-21 22:39:15 Beijing time (10:39 ET, roughly 69 minutes after the open). These are unsettled and are not used to draw settled conclusions.
• All returns are raw price returns, with no dividend adjustment. Note: the S&P 500 ETF went ex-dividend $1.889 on 09-18, so its 09-18 price return is −0.12% while its total return is +0.13%. This piece uses the former.
• Both Korean large caps were verified against two independent sources; closing prices and percentage changes matched exactly.
• This is public-layer analysis containing market observation, public data and method only. It contains no positions, position sizes, share counts, dollar amounts or trades.

Bottom line

On 18 September, the U.S. market priced assets by storage purity. The closer a company sat to pure storage media, the more it rose, while broad indices actually fell. The rank correlation was +0.918. Three sessions later, in the first hour of the 09-21 session, that gradient collapsed to roughly zero (about −0.06; seven snapshots ranged −0.26 to −0.03). In between, the 09-21 Korean close saw a baton change: Samsung overtook SK Hynix. This is not "the storage rally is over." It is "the market has stopped sorting by flash purity" — a very different claim, and the one the data actually supports.

1. September 18: a near-perfect storage-purity gradient

Tier definition (our own editorial construct, not an official classification)

Settled 09-18 returns, sorted descending

What matters here is the ordering, not the magnitude

• From the purest storage media down to the broadest index, returns decline monotonically. The single crossover is the semiconductor ETF (Tier 1, +2.21%) landing among the logic names (Tier 2), which is unsurprising given that the ETF is built from those very constituents - separating them into distinct tiers is a judgment call in our own convention. Apart from that one crossover there is no exception.
• The broadest bucket is negative. The S&P 500 fell that day, which means this was not "a strong tape lifting storage." It was money moving out of something else and into one narrow category.
• Tier means: Tier 4 +7.35%, Tier 3 +3.92%, Tier 1 +2.21%, Tier 2 +2.01%, Tier 0 +0.26%. Tiers 1 and 2 are effectively tied (0.2 points apart); the real separation is between Tiers 4 and 3 on one side and Tier 0 on the other.
• Pure-storage premium over broad indices: +7.09 percentage points.
• Spearman rank correlation (storage purity × return): +0.918. Excluding the AMD outlier discussed below it is +0.922, so the result does not hinge on any single name.

2. The 09-21 Korean close: the baton changed hands

Korea was the only newly settled session on 09-21, and the only independent cross-market check available here.

• On 09-18, Hynix led Samsung by 3.05 points.
• On 09-21, Samsung led Hynix by 4.39 points.
• Across two sessions, relative strength swung 7.44 points and flipped direction completely.

Why this matters. SK Hynix is the share leader in high-bandwidth memory (HBM), while Samsung carries proportionally more NAND flash and commodity DRAM. In other words, on 09-18 the market paid for the HBM end, and on 09-21 it paid for the flash and commodity-memory end — with the 09-18 U.S. flash repricing sitting precisely between those two Korean sessions.

⚠️ The boundary must be stated plainly: this shows that the relative ordering of prices changed. It does not show who was buying or selling. See section 5.

3. The first hour of the 09-21 U.S. session: the same gradient collapsed

Everything below is intraday and unsettled, captured 2026-09-21 22:39:15 Beijing time (about 69 minutes after the open).

Three quantifiable changes

1. The rank correlation fell from +0.918 to roughly −0.06. The gradient is gone.
2. The pure-storage premium over broad indices compressed from +7.09 points to −0.2 points, i.e. essentially to nothing.
3. SanDisk went from leading the Nasdaq-100 by ten points to trailing it by about two (+10.36pp → about −1.8pp, a swing of roughly 12 points).

But the statistic itself is moving, and that has to be disclosed

We took seven snapshots of the same basket on a single convention between 22:35 and 22:39:

What holds: the rank correlation has collapsed to near zero, the pure-storage premium has essentially disappeared, and SanDisk is genuinely trailing the Nasdaq-100 (negative in all seven samples, with a stable magnitude).

What does not hold: "the gradient has inverted to negative." All seven samples are negative, but all sit close to zero, and the spread of the range exceeds its own absolute value — intraday noise alone is enough to explain the sign. This is exactly why intraday data is not used here to draw settled conclusions.

4. Three pieces of evidence against this read

A judgment is only worth as much as the counter-evidence presented alongside it. All three of the following cut against the flash-repricing thesis.

4.1 The top gainer on 09-18 saw money flowing out

In block-order net flows that session, SanDisk — up 10.99% — saw roughly $39.4 million of net outflow, and AMD roughly $53.3 million of outflow, while the two largest net inflows were Micron (about +$156.0 million) and Nvidia (about +$105.0 million). The leaderboard and the money flow point in opposite directions. One reading is that large capital bought the bigger, less volatile end of the complex rather than the highest-beta end — but that is only one reading, and this piece does not adjudicate it.

4.2 The options market did not follow the rally at all

At the 09-18 close, implied-volatility percentiles were on the floor: Micron 4.1, SanDisk 5.7, and both the semiconductor ETF and TSMC at 0.0. For Nvidia and the semiconductor ETF, implied volatility minus trailing 20-day realized volatility came in at −10.55 and −4.67 respectively. Two negative numbers. The options market was pricing a gradual, persistent slope, not a one-off event.

⚠️ Sampling caveat: these percentiles are computed over a self-collected sample of 96 to 125 trading days, not 252. Any description of them as a "one-year low" would be wrong.

4.3 AMD's move is an unexplained outlier

AMD was +8.92% intraday on 09-21, the largest move in the table and unrelated to storage purity. This piece does not attribute that move to any news item, because no verified primary source was available. We did verify that excluding AMD leaves the 09-18 correlation at +0.922 and does not change the 09-21 conclusion, so the main finding does not depend on this outlier.

5. Two questions this piece cannot answer

5.1 Who bought Samsung's 4.98%? Unknown.

Korean single-stock foreign investor flow data is 12 sessions behind (latest available: 09-09). This is a collection gap, not evidence that foreign investors were inactive. Foreign buying, domestic institutional buying, and no dominant buyer are therefore all three impossible to rule out. Until that data is restored, this piece makes no claim about who was buying or selling.

5.2 Why didn't Hynix follow? There is a primary-source lead, but it is unresolved.

On 09-18, SK Hynix filed a reply to a Korea Exchange rumor inquiry (document number 20260918800583). The conclusion was "undetermined."

Under Korea Exchange disclosure convention, this is not a denial — a denial requires the explicit wording "no basis in fact." The reply carries a hard re-disclosure deadline of 2026-12-17, and it is already the second deferral (the first inquiry was dated 08-21). The inquiry relates to reports of a plant in Japan.

This is currently the only pending item with a hard date attached. It is left open here, and no directional conclusion is drawn from it.

6. How to test this at tonight's close: three falsification bands

Intraday does not count. Tonight's 09-21 U.S. close does. The bands are fixed in advance and will not be moved after the fact.

Second cross-market leg (09-22 Korean close):

• Hynix outperforms Samsung → the 09-21 baton change was single-session noise and is void.
• Samsung leads by more than 2 points again → two consecutive sessions in the same direction, and only then may "a NAND versus HBM divergence" enter a formal judgment.
• The gap falls between 0 and 2 points → report "no conclusion."

7. Method and conventions

• The tiering is our own editorial construct, based on how much of each company's revenue sits in storage media. It is not an official industry classification, and alternative tierings may produce different coefficients.
• Return convention: raw price return (close ÷ prior close − 1), with no dividend adjustment. Intraday 09-21 returns use the 09-18 raw close as the denominator, so they are directly comparable to the settled 09-18 returns.
• Why this is emphasized: the S&P 500 ETF went ex-dividend $1.889 on 09-18. On a total-return basis its 09-18 reading becomes +0.13% instead of −0.12%, which would erase the fact that the broad index closed red that day. Because the dividend convention changes the qualitative conclusion, it must be declared.
• Sample size: eleven names is a small sample. At this size the confidence interval around a rank correlation is wide, so the coefficient is used here as a descriptive statistic and no significance claim is made.
• Intraday data was taken from a real-time quote interface and cross-checked for directional agreement against a second independent source. Settled data was verified against an independent historical source.
• Korean data: index and single-stock readings were compared across two sources; both large caps matched exactly on closing price and percentage change.

Frequently asked questions

Is the storage rally over?
This piece makes no such claim. What the data supports is narrower: the market has stopped sorting by flash purity. An axis losing explanatory power and a rally ending are two different things. Intraday on 09-21, Micron was still up 2.48% and the semiconductor ETF up 2.46% — the difference is that broad indices rose just as much or more.

Does Samsung overtaking SK Hynix mean HBM is weakening?
That inference is not available yet. Only one settled session points that way, and the session before it pointed the opposite way. Under the criteria in section 6, two consecutive sessions in the same direction are required before this enters a formal judgment.

Why not use P/E ratios to judge whether storage stocks are expensive?
Storage is a deeply cyclical industry, and P/E ratios are at their highest at the earnings trough, so percentile rankings invert. In addition, the companies discussed here have different fiscal year ends, which makes a direct cross-sectional comparison invalid on its own terms.

Public-layer market observation based on settled data through the 2026-09-18 U.S. close and the 2026-09-21 Korean close, plus intraday U.S. readings captured at 2026-09-21 22:39:15 Beijing time. Not investment advice.