Foreign Flow Watch: A Fourth Buyer Absorbed the $5.1bn Foreigners Sold in Samsung and SK hynix — Week of August 28, 2026

Foreigners sold $5.1bn of Samsung Electronics and SK hynix in five sessions. Someone bought $5.76bn of them. It was not retail, it was not the institutions, and it was not foreign money coming back. It was a fourth category — and two Korean-language filings, posted the week before, say exactly who it was and exactly which day it would start.

This is Foreign Flow Watch, our weekly read of where global capital actually moved in Korean equities. Every figure below is computed from settled exchange data for 2,551 KOSPI and KOSDAQ common stocks. Week of 24–28 August 2026 — 5 sessions. Previous issue: Issue #6, week of 21 August. Next issue: Issue #8, week of 4 September.

📉 The two boards traded places again

Last week the KOSPI was the stronger board and the KOSDAQ fell. This week that reversed, and the two boards finished 6.34 percentage points apart. Measured consistently — from the previous Friday’s close, the convention we settle on in the next section — Korea’s two exchanges have finished more than six points apart in three of the last four weeks, and have moved in outright opposite directions in two of them.

The path matters more than the endpoint here:

📐 Which Monday? Why this week’s KOSPI has two different numbers

The KOSPI fell 1.79% this week if you measure from the previous Friday’s close, and rose 1.37% if you measure from Monday’s close — and this series has published both conventions without saying so. We found this while grading last week’s issue, so we are printing it before anything else.

The problem is simple. Our flow window is five sessions, Monday through Friday. The return over those same five sessions runs from the previous Friday’s close to this Friday’s close. If you instead start the clock at Monday’s close, you throw away Monday’s return entirely. This week Monday was −3.12%, the largest single move of the week, so discarding it flips the sign of the answer.

From this issue forward the weekly index change is measured from the previous Friday’s close — the standard weekly window for this series — because that is the only start date that matches the five sessions of flow data sitting beside it. Three corrections follow from that, and all three go against us:

• Issue #6’s KOSPI was not up 0.63%. On the standard weekly window it was down 0.93%, and the KOSDAQ was down 7.25%, not 3.87%. Both boards fell; the KOSDAQ simply fell far harder. The sentence “the KOSPI closed green” does not survive, and neither does the framing of a rising large board against a falling small one.
• Issue #6’s opening line was right and its table was wrong — which is the opposite of what it looks like at first glance. The lede said the two boards finished “six percentage points apart”, and on the standard window the gap is 6.32 points. The table beside it printed +0.63% and −3.87%, which are 4.50 points apart. The two were computed from different start dates within the same issue: the lede from the prior Friday, the table from Monday. The number that reads like the error is the one that was correct.
• The correction also cancels one of Issue #6’s own scorecard calls. It marked the condition “the pair keeps being sold while the index keeps rising” as Triggered, on the basis that the KOSPI rose 0.63%. On the standard weekly window the KOSPI fell 0.93%, so the index was not rising and that condition should never have been marked triggered. Issue #6 graded itself more harshly than the data warranted. We are leaving its reasoning visible rather than deleting it, and we count the reversal here.

Issue #4’s reading is unaffected: it already used the standard weekly window, and on the Monday-close convention its central finding would have been invisible, since the index would have looked flat at +0.02%. See our corrections policy.

🧮 Two stocks were 196% of the KOSPI’s decline

Aggregate market value of the 829 KOSPI common stocks in our reference set fell from $3,991bn to $3,917bn, a loss of $73.6bn, or 1.84% — close to the index’s 1.79% and a useful check that the two agree.

The two memory names lost $144.5bn between them. The index lost $73.6bn. The pair therefore accounts for 196% of the decline — everything else, in aggregate, pushed the other way. Strip those two out and the remaining 827 KOSPI companies gained 3.77%, which is closer to the KOSDAQ’s +4.55% than to the headline.

Samsung and SK hynix were 52.8% of KOSPI market value on 21 August and 50.1% on 28 August. Issue #4 made this same measurement three weeks ago and got 140%. It is now the third time this month that a KOSPI print has described two companies rather than a market. That is not a market call; it is a warning about the number itself. If you are reading the KOSPI as a gauge of Korean corporate health, roughly half of it is a bet on memory pricing — which we set out in Korea’s chip supercycle.

🌊 Who sold, and who did not

Korea’s exchange reports flows in three named categories: individuals, foreigners, and institutions. Everything else — chiefly corporations trading their own or others’ shares — falls into a residual that our data does not name. Normally that residual is small enough to ignore. This week it was the largest buyer in the market.

Method: each session’s net share count is valued at that session’s own close, then converted at that session’s own USD/KRW close. The reference rate for market-value figures is ₩1,380.45/$ as of 28 August 2026, from FinanceDataReader. These are approximations from settled net positions, not execution prices. The four rows sum to zero by construction, which is how we recover the residual — we do not receive it directly.

Retail did not repeat last week’s absorption. In Issue #6 domestic individuals were a $1.84bn net buyer. This week they were a $0.57bn net seller. Institutions were sellers on both boards. The only meaningful net buying in Korea this week came from foreigners on the KOSDAQ and from the residual on the KOSPI.

🏦 What is the fourth buyer?

The residual is almost entirely two companies buying their own shares on the open market, under buyback programmes filed with Korea’s DART disclosure system on 19 and 21 August 2026. We cannot identify the buyer directly — our data reports a category, not a name — but the alignment is close enough that we are printing it, along with the tests we ran to try to break it.

First, the residual is not a market-wide phenomenon. Of the $5,768M it bought on the KOSPI, $5,762M was Samsung Electronics and SK hynix. Across the other 827 KOSPI names the residual netted roughly $11M for the week — the same near-zero it has been all month. The third-largest residual buy in the entire market was Krafton (259960) at $14.9M, some 263 times smaller than SK hynix’s $3.92bn.

Second, it started on a specific day in each name, and the two days are different:

Net shares bought by the unnamed residual category, from settled exchange flow data. Negative values mean the residual was a net seller.

SK hynix’s residual switches on at 20 August and then holds a band of 635,000 to 651,000 shares a day — a 2.5% spread across six sessions. Samsung’s stays negative through 21 August and switches on at 24 August, then holds 1.84 to 2.03 million shares a day. Two different companies, two different start dates, both stable to within a few percent thereafter. That is the signature of a scheduled programme, not of discretionary trading.

Third, those two dates are written down in advance. SK hynix filed a treasury-share acquisition decision with DART on 19 August 2026 naming an acquisition period that begins 20 August. Samsung Electronics filed its own on 21 August naming a period that begins 24 August. Both specify on-market purchase. Both carry a daily order-quantity cap, and the observed daily volumes sit comfortably inside them.

Primary sources: SK hynix treasury-share acquisition decision, rcpNo 20260819000254 and its companion share cancellation decision, rcpNo 20260819800340; Samsung Electronics treasury-share acquisition decision, rcpNo 20260821000616. All three are Korean-language filings on DART, Korea’s statutory disclosure system.

One cross-check on those documents. Each filing derives its daily cap partly from “1% of shares outstanding”, and states that figure: 58,462,786 for Samsung and 7,304,923 for SK hynix. Multiply by 100 and you get 5,846,278,600 and 730,492,300 shares — against 5,846,278,608 and 730,492,365 in settled listing data from the Korea Exchange (KRX). The filings and the exchange agree to within rounding, which is how we are confident we are reading the right line of the right document.

⚖️ Does a $40bn buyback mean the same thing as a $15bn one?

No — and here the difference is the whole point, because only SK hynix’s programme reduces the share count. The two filings look alike in an English summary and are opposite in effect.

SK hynix states the purpose as acquisition for cancellation, and filed a separate cancellation decision the same day. Shares bought under that programme are retired: 24,070,000 shares, or 3.30% of the company, permanently gone. Every remaining shareholder’s claim on earnings rises by that fraction.

Samsung Electronics states the purpose as employee share-based compensation — performance incentives and special performance bonuses. Those 53,285,968 shares are bought from the market and then handed to employees. The share count does not fall. The stock is transferred from public float to staff, and the buying pressure is temporary by design.

Both are real cash leaving the company. Only one is a return to shareholders in the sense the Value-Up programme is trying to encourage. This distinction — buyback versus buyback-and-cancel — is the single most useful thing a foreign investor can learn to read off a Korean disclosure, and it is the reason we go to the filing rather than the summary. We made the same point about Celltrion, which cancelled 4% of itself and still ended the year with more shares outstanding.

One more line from the filings is worth having. Samsung’s programme covers common shares only; the preferred line is explicitly excluded. Over the same five sessions Samsung’s common stock fell 8.70% to $186.17 (₩257,000) and the preferred fell 9.76% to $135.32 (₩186,800), with foreigners selling $720.4M of the preferred. We are not claiming the exclusion caused the gap — one week and one percentage point is far too little for that — but the reader who wants to test the buyback’s price effect now knows where the control group is. Our full breakdown of both listings is in Samsung and SK hynix stock prices in USD.

⚠️ Where flow and price disagreed this week

We print these every week, because a table of “foreigners bought it and it went up” selected after the fact is not evidence of anything. Here is the same week’s data where the relationship fails.

The strongest single example is not in the table, because its foreign flow was almost nil. Doosan Enerbility (034020) added $6.9bn of market value, the largest gain on the KOSPI, on foreign net selling of $4.6M — a flow so small it would not register on any ranking. Whatever moved that stock 20.33% in five sessions, it was not foreign money.

SK Square deserves a note of its own. Foreigners were its fifth-largest net buy in the market at $75.9M, and it fell 8.46%, losing $9.1bn of market value. SK Square’s principal asset is its stake in SK hynix, so a week in which hynix fell 4.45% is a week in which the holding company’s net asset value fell too — the discount mechanics we cover in our SK Square valuation. That is a plausible story. It is also a story we constructed after seeing the number, which is precisely why it sits in the disagreement table rather than in the argument.

🧾 Scorecard: how Issue #6 held up

Issue #6 published four items under “prove this issue wrong”: three testable conditions and one self-flagged weak number. None of the three conditions fired — but read the second column before scoring that as a win, because every one of them failed on its second clause, not its first. The fourth item turned out to be a caveat we did not need.

Three tests that could not run is a worse outcome than a clean failure, because a falsifier with a second clause we do not control is not much of a falsifier. Issue #6 also stands corrected on its index convention, as set out above — and that correction runs in both directions. It cancels one of Issue #6’s own gradings: the condition marked Triggered because “the KOSPI rose 0.63%” should not have been, because on the standard weekly window the KOSPI fell 0.93%. Issue #6 scored itself harder than the data justified, and its “six percentage points” lede — the line that looked like the mistake — was the part that was right. Our record grading ourselves now runs: Issue #4 scored Issue #3 at two failures out of three, Issue #5 scored Issue #4 at three holds and one open, and Issue #6 scored Issue #5 at two failures out of three. We print the bad weeks at the same size as the good ones.

🧩 Prove this issue wrong

• If the residual buying in Samsung or SK hynix stops or halves while the filed acquisition window is still open, then the residual was never mostly the buyback and our identification is wrong. Both windows run to mid-November, so this is testable weekly for twelve more weeks.
• If either company’s next quarterly treasury-share report shows purchases materially below what we attributed to it, the arithmetic here is an upper bound that we leaned on too hard. Korean issuers must report execution; that filing, not our residual, is the authority.
• If the residual appears in a third large-cap name with no corresponding DART filing, then the pattern is a data artefact of our flow source rather than a description of corporate behaviour.
• If restating Issues #4 through #6 on the standard weekly window changes any of their conclusions, and not merely their headline percentages, then this week’s methodology fix is a bigger correction than we have presented it as. We restated #6 in full and it cost us a headline, a framing and one of its own scorecard grades; #4 was already on this window. We have not re-derived #5 in full.
• Our weakest number this week is the USD/KRW date alignment. Our FX source labels its rows on a trading calendar that is offset from Korea’s, so we map each row forward one business day. That mapping reproduces the ₩1,390.79 rate Issue #6 published for 21 August, which is our only external check on it. The week’s rates span ₩1,380.45 to ₩1,384.98, so a one-day error moves any dollar figure here by less than 0.35% — but it is an inference, not a reading.

📚 Lingo Check

🎯 Why it matters for K-Export Stars

A single KOSPI print told you Korea fell 1.79% this week. Underneath it, the other 827 KOSPI companies rose 3.77%, the KOSDAQ rose 4.55%, foreigners pulled $4.91bn out, and the largest buyer in the country was two issuers executing programmes they had disclosed in Korean five and three days earlier.

None of that requires a view on memory prices. It requires reading a filing. Both documents were public before the buying started, both name the exact start date, and both state a purpose that determines whether the share count falls — and one of them says it does not. An investor working from English-language coverage of “Samsung and SK hynix announce buybacks” would have had the direction right and the mechanism backwards.

That is the whole thesis of this site, and this week it happens to be worth $39.85bn of announced buying. For how these flows are constructed and why they mislead as often as they inform, start with how foreign and institutional flows move Korean stocks.

Every figure here is reproducible from our open dataset — 2,551 KOSPI and KOSDAQ common stocks, updated every trading day, CC BY 4.0, DOI 10.5281/zenodo.21833374. Index levels and closing prices are settled data from the Korea Exchange (KRX). Buyback terms are quoted from the DART filings linked above. This issue makes no causal claim about flows; where flow and price disagree, we print the disagreement.

Disclosure of interest. The author holds positions in Samsung Electronics (005930) and SK hynix (000660), the two stocks this issue is largely about, and in other Korean equities not discussed here. No position was opened or closed in any stock named above in connection with this article.

Disclaimer: This article is for informational and educational purposes only and is not investment advice, a recommendation, or an offer to buy or sell any security. Figures are as of 28 August 2026 and go stale quickly. Do your own research and consult a licensed adviser before investing.