Did Foreign Investors Buy Korea or Just Buy Memory on September 7, 2026? A KRW-Weighted Decomposition Shows 97% of Buying Went to Three Names

Data basis (read this first)

• Korean figures: settled closing values from the Korea Exchange (KRX) session of 2026-09-07. This is the only new trading session in this piece.
• U.S. figures: U.S. markets were closed on 2026-09-07 for Labor Day. Every U.S. number below is the settled close of 2026-09-04 and is labeled as such at each point of use.
• Hong Kong: not cited. No settled reading entered our data layer for this date.
• FX: KRW/USD is a live rate, as-of 2026-09-07. USD figures are converted for order-of-magnitude reference only; the KRW originals are authoritative.
• This is public-layer research containing only public market data, methodology, and opinion. It contains no position, sizing, or trade information.

The one-sentence conclusion

In the Korean session of 2026-09-07, counting names and counting money give opposite answers. The number of names foreign investors net-bought expanded from 2 to 6 (which looks like broad re-risking across Korea), but 97.11% of the buying value landed on just three memory-chain names (which means it was still a memory trade). Both readings are correct — they simply answer different questions. To judge positioning, use weight. Breadth only tells you about sentiment.

What happened: breadth genuinely flipped

Across the eight Korean names we track, comparing the previous session (09-04) with this one (09-07):

On this table alone the conclusion looks straightforward: foreign investors flipped from broad selling to broad buying and scaled the amount by nearly 5x — a textbook market-wide re-risking.

The supporting evidence looks strong too, because it was not only semiconductors that flipped. Hyundai Motor (−60,899 → +21,408 shares), Kia (−69,915 → +46,829) and LG Electronics (−89,824 → +248,325) — three names with no HBM exposure — all flipped from net-sold to net-bought. If this were purely a memory trade, they had no reason to flip in the same session.

But the picture inverts once you weight by money

Multiply each name's foreign net-bought share count by that session's closing price to get the net-bought value, then express it as a share of gross buying (the six net-bought names only, totaling KRW 2.3413tn):

Three ratios matter:

• SK Hynix + Samsung Electronics = 96.23% of gross buying (98.05% of net buying)
• Adding Hanmi Semiconductor (HBM stacking/bonding equipment), the three memory-chain names = 97.11% of gross buying
• The remaining five non-memory names = 2.89% of gross buying combined

Put differently: the four additional names foreign investors bought contribute almost nothing in value terms. LG Electronics, Hyundai and Kia together received KRW 67.6bn — just 4.9% of what SK Hynix alone received (KRW 1,376.5bn).

Why the distinction matters

Breadth and weight are two different statistics answering two different questions, and market commentary routinely conflates them:

• Breadth answers "how many names were touched" — it is sensitive to sentiment and participation. The 6-buy / 2-sell flip genuinely does indicate that foreign attitude toward Korea improved at the margin. That layer is real.
• Weight answers "where did the money actually go" — it is sensitive to positioning and risk exposure. A 97.11% concentration in the memory chain means the exposure actually added is still a single theme.

The consequence of conflating them is concrete. If you read this session by breadth as "broad Korean re-risking," the natural next inference is that non-memory sectors should follow higher. But the value structure shows the non-memory names received amounts small enough to ignore — their gains are more likely passive drift than an independent move with capital behind it.

The price structure agrees with the value structure, not with breadth. Rank the eight names by session move and there is a 2.09 percentage-point gap between the top four (+8.26%, +6.20%, +5.68%, +4.57%) and the bottom four (+2.48%, +1.50%, +1.12%, −1.28%) — specifically between Hanmi Semiconductor at +4.57% and Hyundai Motor at +2.48%. That is stratification, not a broad rally.

Three tail details

1. The two names still being sold are both batteries. LG Energy Solution (−4,458 shares) and Samsung SDI (−77,200 shares) are the only two net-sold names, and Samsung SDI was the only one of the eight to close lower (−1.28%), with its foreign net selling widening versus the prior session (−57,443 shares). "Market-wide re-risking" does not hold in the battery bucket — this is the most direct counterexample to an overreaching breadth narrative.

2. The equipment leg is shrinking, not growing. Hanmi Semiconductor saw foreign net buying of 353,063 shares in the prior session, falling to 86,155 this session — a 75.6% contraction. It was one of only two net-bought names on 09-04, yet it now accounts for 0.88% of gross buying. Capital is making choices within the memory chain too, concentrating into the upstream device makers (Hynix, Samsung) rather than spreading evenly along it.

3. The slow variable agrees. Foreign ownership ratios — far less sensitive to single-session noise — moved the same way: SK Hynix 50.50% → 50.61%, Samsung Electronics 46.73% → 46.78%, LG Electronics 29.46% → 29.62%. Samsung SDI fell from 25.62% to 25.53%. Direction is consistent with the value structure.

U.S. markets: no new evidence this date

2026-09-07 was Labor Day in the United States. U.S. markets were closed and no new U.S. price or volume evidence exists for this date. The last settled session was 2026-09-04:

One cell deserves separate mention: NVIDIA's option-implied volatility (33.21) sits 11.59 points below its own trailing 20-day realized volatility (44.80). The options market is pricing NVIDIA's future movement well below the movement it just actually delivered. SanDisk (−2.74) is a second, milder instance of the same sign. The other seven rows carry positive premiums.

Methodological warning (must be cited alongside the numbers): the IV percentiles behind this table (NVIDIA 0.9, Broadcom and both broad ETFs at 0.0, SanDisk 5.3, Micron 4.4) are computed on a self-collected sample of n=86–115 trading days, not the industry-standard 252. Percentiles computed on different windows are not directly comparable to external data sources. We use them here only for within-table ranking, never for an absolute high/low claim.

Falsification criteria

The core claim of this piece is: this session was theme-level accumulation in memory, not market-wide Korean re-risking. We track it with a continuous measure rather than two threshold endpoints, so there is no undefined middle.

Define concentration C = foreign net-bought value of the three memory-chain names (000660 + 005930 + 042700) ÷ total foreign gross buying that session. This session, C = 97.11%.

Over the next three Korean trading sessions:

Independent second leg: if LG Energy Solution and Samsung SDI both flip to foreign net buying, the "batteries are still being reduced" counterexample disappears and the breadth narrative gains real support — the strength of this piece's conclusion should be marked down even if C remains high.

Known conditions that distort this criterion: the Korea Exchange revises same-day data after the close (our readings freeze at the close and revisions are not backfilled); passive flows on MSCI/FTSE rebalance dates distort single-session structure in one shot; and Samsung Electronics and SK Hynix carry such large index weights that any index-level flow mechanically raises C — in that case a high C does not indicate active thematic selection. Confirm none of these events applies to the session before citing this criterion.

Method: how to reproduce this yourself

1. Take the foreign net-bought share count per name from the Korea Exchange daily investor-type trading data (publicly available).
2. Multiply by the closing price of the same session to get net-bought value. It must be the same session's close; crossing sessions injects price drift.
3. Use gross buying (sum over net-bought names only) as the denominator, not net. Using net lets the sellers cancel part of the denominator and inflates the concentration figure. We give both here: 98.95% of net, 97.11% of gross. Cite the latter — it is the more conservative one.
4. Present it alongside the name-count measure. Reporting only one of the two is precisely the error this piece is about.

Frequently asked questions

Q: How much Korean stock did foreign investors buy on September 7, 2026?
A: Across the eight names we track, foreign net buying was KRW 2.298tn (~USD 1.71bn) at closing prices, 4.94x the prior session's KRW 0.465tn. Note this is an eight-name basis, not a whole-market figure.

Q: Does this mean foreign investors are bullish on Korea as a whole?
A: By name count, yes (6 bought vs 2 sold). By value, no (97.11% concentrated in three memory-chain names). The value basis better reflects the risk exposure actually added.

Q: How much of it was SK Hynix and Samsung Electronics?
A: 58.79% and 37.44% of gross foreign buying respectively — 96.23% combined.

Q: Which names were still being sold by foreign investors?
A: LG Energy Solution (−4,458 shares) and Samsung SDI (−77,200 shares), both batteries. Samsung SDI's net selling widened versus the prior session and it was the only one of the eight to close lower (−1.28%).

Q: How did U.S. markets perform on September 7, 2026?
A: U.S. markets were closed for Labor Day; there was no trading. The last settled session was September 4, when SanDisk rose 11.90%, Micron 6.10%, the SMH semis ETF 2.61%, while the SPY S&P 500 ETF fell 0.39%.

Produced by the AlphaGBM research team from public market data, published 2026-09-07 22:30 Beijing time. All Korean figures are settled KRX closes for 2026-09-07; all U.S. figures are settled closes for 2026-09-04. For research and educational purposes only. Not investment advice. Contains no position or trade information. Verify all data independently before acting.