BS Report

South Korea's AI Bubble Just Popped

South Korea's AI Bubble Just Popped — Andrei Jikh, YouTube, 2026-07-20 · 2,719,326 views · 70,700 likes · 3.33M subscribers

2026-08-02 · bullshit-detector 0.13.0

4/10
Mostly fine

Mostly fine

Tally: 60 claims extracted, 44 individually source-checked — 35 confirmed, 5 plausible, 7 misleading, 3 false. 6 unverifiable; 2 not checked; 2 not rateable.

Ambiguous: 2 claims dropped before verification — that hyperscaler spending "happens to be the revenue of Nvidia, of Micron, and of the whole AI trade" (a sum of three revenues, each of them separately, or a notional sector aggregate — the readings settle differently), and "they're saying that the first hyperscaler to cut back on AI spending will be rewarded by the market", where they is never identified.

What it says (neutral summary)

South Korea's KOSPI, the world's best-performing market of 2026, collapsed roughly 25% in three weeks after peaking on 22 June, triggering mass margin calls and forced liquidations among the country's 14 million retail investors, who had concentrated leveraged bets on two AI-memory stocks that make up more than half the index. The video argues this was caused by American chip-sector nerves, and that the US is running the same mechanism more slowly: record margin debt at 4.5% of GDP, a top-heavy S&P 500, and an AI capex loop in which four hyperscalers' spending is the entire sector's revenue. It closes with two scenarios — Korea stays contained, or Korea was the first domino — and says the number to watch is hyperscaler capex guidance.

Load-bearing claims

The ones the thesis dies without. Verify all of them.

#Claim (with timestamp)TypeVerdictEvidence
1[01:04] [After peaking on 22 June 2026] "in just 21 days, everything started to collapse. The KOSPI dropped 25%"factual✅ confirmedExact. Record close 9,114.55 on 2026-06-22 → 6,806.93 on 2026-07-13 = −25.3%, and 22 Jun → 13 Jul is 21 days to the day. Daily closes, KOSPI (^KS11) (tier 1, exchange price data)
2[00:00] [In the KOSPI] "just two companies alone, Samsung and SK Hynix""represent over 56% of their whole stock market"factual✅ confirmed3 URLs → 3 origins: crossed 50.44% of KOSPI market cap in late May 2026 (Seoul Economic Daily), 50%+ of KOSPI 200 for the first time (Asia Business Daily), and a record ~60% by 30 June 2026
3[00:00] The KOSPI "gone up almost 200% in the last 12 months" [measured to its peak three weeks before publication]factual✅ confirmedBasis matters and the wording fixes it — the sentence opens "just 3 weeks ago". 12 months to the 22 Jun peak: 3,188.07 (18 Jul 2025) → 9,114.55 = +185.9%. Measured instead to the day of publication (20 Jul close 6,516.27) it is +104.4%. Bloomberg used the same peak basis: "the Kospi surging close to 200% over the last 12 months". Price data
4[01:04] "1.2 million accounts" [in South Korea] "got hit with something called margin call thresholds"factual✅ confirmed3 URLs → 2 origins: more than 1.2 million retail leveraged credit accounts triggered margin calls as of 13 July 2026
5[01:04] "over 3 trillion won in investments were liquidated"factual✅ confirmedJuly forced liquidations reached 3.442 trillion KRW, with a single-day peak of 1.422 trillion won on 9 July
6[09:37] On the day Koreans call Black Tuesday, "the KOSPI went down more than 10% in just a day"factual🟠 misleadingThe pre-publication Black Tuesday was 23 June 2026: −9.99% — and the close was the intraday low, so it never went below −10% at any point in the session. Price data; Asian markets' "Black Tuesday", KOSPI −9.99%, Samsung and SK Hynix both −12%. A KOSPI day worse than −10% did arrive — −10.84% on 28 July, eight days after this video was published
7[10:40] "about 2% of Korean margin accounts get force liquidated. But during this crash, that number went higher than 10%, which is about five times the normal rate"factual✅ confirmedAbove 10% in the week of 10 July against a 2.1% average over the prior six months. 10 ÷ 2.1 = 4.8×, i.e. "about five times"
8[03:12] Of South Korea's population, "14 million of them, or one in four, are what's called retail investors"factual✅ confirmed2 URLs → 2 origins: 14 million retail investors (Bloomberg; also the figure cited when Korea scrapped its capital-gains tax in 2024). 14 ÷ 51.7 = 27%, so "one in four" holds
9[06:26] AI data centres need memory chips and "the two companies that dominate that market are Samsung and SK Hynix"factual✅ confirmedQ1 2026: Samsung 38% + SK Hynix 29% = ~67% of global DRAM; SK Hynix 58% + Samsung 21% = ~79% of HBM. Counterpoint DRAM/HBM share
10[06:26] Microsoft, Google, Amazon and Meta "have spent $376 billion on AI infrastructure in 2025"factual✅ confirmed2 URLs → 2 origins: combined Big Four 2025 capex over $378bn. (A wider capex definition puts 2025 at $410bn — his figure sits on the AI-infrastructure narrower one)
11[06:26] "this year in 2026, they're on track to spend 725 billion"factual✅ confirmed$725bn planned 2026 capex, +77% year on year — Amazon ~$200bn, Microsoft ~$190bn, Google $175–185bn, Meta $115–135bn
12[13:52] "As of June 2026, US margin debt just hit roughly 4 and 1/2% of GDP. That is the highest level ever recorded in American history"factual✅ confirmedActually 4.71% — he rounds down. FINRA margin debt / GDP, June 2026; $1.53tn, +7.9% in June, +51.5% y/y, a record (published 20 July, the day of the video, matching his "released this past week")
13[13:52] It is "higher than the dot-com bubble, higher than 2007, higher than even the 2021 meme stock everything bubble, and it's higher by a lot"factual✅ confirmedPrior cyclical peaks: 2.8% (2000), 2.6% (2007), 3.8% (2021) against 4.71% now — 24% above the 2021 peak and 68% above 2000. Historical series
14[13:52] "every time US margin debt went to an extreme, a crash happened almost right after, every single time"factual🟠 misleadingThe named peaks are real — margin debt did top out near the 2000, 2007 and 2021 highs — but the framing inverts what the indicator is. Margin debt is a coincident measure of risk appetite, not a timing signal: it moves with the market, and "can stay at record levels for a long time while markets rise". The chart marks the extremes that were followed by crashes; the record months that weren't get no arrow
15[16:01] "The top 10 stocks are 36% of the S&P 500", a share "which historically is a very high concentration"factual✅ confirmed36.4% as of 29 May 2026, against 23% in 2000
16[18:12] Of AI compute, OpenAI and Anthropic "make up 70 to 80% of all the AI compute demand"factual❌ falseOff by roughly 3.5×. Epoch AI: OpenAI, Anthropic and xAI together held ~21% of global AI compute (range 20–30%) at end-2025, operating under 4 million H100-equivalents. Most compute sits with cloud providers serving everyone else. This row carries the "pyramid structure" argument — if two labs are a fifth of demand rather than three-quarters, their unprofitability does not sit under the whole stack
17[17:06] Jim Chanos says hyperscaler returns "gone from as a group 40% a year and a half ago to about 20% today", heading toward 10%factual✅ confirmedChanos did say this. "hyperscaler incremental return on invested capital is plunging from 40% to 20% and sliding toward 10%". Note what is confirmed: that a named short-seller estimates this, not that the estimate is audited
18[18:12] "In 2026, IPOs have come to market at over 12% of GDP in post-IPO market cap. Now, the previous record was the peak of the dot-com bubble at just 5%"factual🟡 plausibleSearched three angles — the stat itself, Goldman's 2026-vs-1999 IPO work, and the 2026 mega-listings — and the series was never located. What is checkable: SpaceX listed at a ~$2.43tn market cap, roughly 8% of US GDP on its own, so 12% across all 2026 listings is not absurd. But Goldman's read of 2026 issuance is that it "still falls short of bubble territory", and the two largest remaining listings (Anthropic, OpenAI) were expected in Q4 2026 — after this video
19[08:32] The US popped Korea's bubble: "In late June, chip stocks in the US started selling off. That's because investors were getting nervous ahead of earnings from a very important company for the AI story, which is Micron."factual🟠 misleadingThe causation runs the other way. On 23 June Micron fell 13% because South Korea's Financial Supervisory Service warned about the Samsung and SK Hynix leveraged ETFs, alongside a report that OpenAI might delay its IPO. Micron then reported the next day and the stock rose on the results. Korea's own regulator lit the fuse; the video reverses it to make the US the cause
20[02:10] What happened in Korea "might actually be a preview of what's about to come to the US"predictionHedged, and hedged again at [22:27] with two explicit scenarios. Not rateable; judge it on the premises above

Incidental claims

Supporting detail. Wrong here is embarrassing, not fatal.

#Claim (with timestamp)TypeVerdictEvidence
21[00:00] "just 3 weeks ago, South Korea had the best-performing stock market in the world"factual✅ confirmed2 URLs → 2 origins: CNBC calls it "the world's best-performing stock market this year"; still +41.5% YTD in USD terms and best-performing major market in late July
22[00:00] Over the same 12 months, other stock markets "like the US's S&P 500, which only made 24%"factual✅ confirmedNo search of its own — computed from the same exchange price series as claim 1, on the basis claim 3 fixes (12 months to 22 Jun 2026): 5,967.84 → 7,472.79 = +25.2%
23[00:00] "At one point this year, one of their companies, Samsung, was up over 500%"factual🟠 misleadingNot on the basis he states. Calendar 2026 to Samsung's 18 Jun peak: 119,900 → 362,500 = +202%. The 500% figure needs a trailing-12-month basis (Jun 2025 → Jun 2026 peak = +506%) or trough-to-peak (+450%), neither of which is "this year". No search of its own — computed from the same exchange price series as claim 1
24[00:00] "SK Hynix was up over 1,000%" [at one point this year]factual🟠 misleadingSame rebasing. Calendar 2026 to the 22 Jun peak: 651,000 → 2,919,000 = +348%. Trailing 12 months to that peak = +1,025%; 52-week trough to peak = +1,091%. No search of its own — computed from the same exchange price series as claim 1
25[02:10] "320,000 accounts were just wiped out, some of them overnight"factual✅ confirmed320,000–360,000 accounts forcibly liquidated
26[01:04] 1.2 million margin-called accounts "works out to be one in every 30 people in the country"factual❌ falseRests on claims 4 and 27. 51.7m ÷ 1.2m = one in 43, not one in 30 — "one in 30" would need 1.7m accounts. (And accounts are not people; one investor can hold several)
27[03:12] "in South Korea, there's about 51 million people"factual✅ confirmed~51.7 million; the 14m-of-51m ratio is the standard framing in coverage
28[02:10] "the president of South Korea held an emergency intervention for their stock market"factual✅ confirmedPresident Lee Jae-myung chaired meetings on two consecutive days, 15–16 July, demanding market-stabilisation measures
29[12:45] The president was "the same person who spent a year encouraging people to get into the stock market"factual✅ confirmedLee campaigned on a "Kospi 5,000 era" and described himself as a seasoned retail investor; retail participation surged under his market-reform drive
30[05:20] "even South Korea's government that allowed them this said" it should not have permitted the leveraged productsfactual✅ confirmedThe FSS governor admitted regret over approving single-stock leveraged ETFs, and Lee demanded tighter regulation of them
31[03:12] "housing in South Korea has become so expensive that for a lot of young people, owning a home stopped being a realistic goal"factual✅ confirmed2 URLs → 2 origins: homeownership among under-39s at a record-low 27.7%; Seoul price-to-income ratio 13.9 years against 6.3 nationwide
32[03:12] Korean retail investors — "they call themselves ants because individually, one ant is small"factual✅ confirmed개미 / gaemi. Reuters and others use the term as the standard label
33[05:20] "At the peak, Koreans invested about 10 trillion won into leveraged funds betting on just one company."factual🟡 plausibleRight order of magnitude, not separately published for a single name. Samsung + SK Hynix single-stock leveraged ETFs held ~14tn won AUM by mid-June and drew 23tn won of inflows since launch; retail net purchases alone were 14tn won
34[08:32] Micron "is America's biggest chip maker"factual❌ falseNvidia is, by an enormous margin — ~$4.85tn market cap in July 2026, the largest semiconductor company in the US by both market cap and revenue. Micron ranks around 6th among US chip firms. He is right that it is America's biggest memory maker, which is the claim the argument needed
35[09:37] "in one session, the KOSPI dropped something like 4.6%"factual🟡 plausibleHedged, and no session in the window closed at exactly −4.6%. Nearest: −4.91% (7 Jul), −4.52% (10 Jun), −4.46% (20 Jul). No search of its own — computed from the same exchange price series as claim 1
36[09:37] "Samsung and SK Hynix both went down more than 9% in just one day"factual✅ confirmedTwice over. 23 Jun: Samsung −12.31%, SK Hynix −12.47%. 13 Jul: −10.70% and −15.37%. No search of its own — computed from the same exchange price series as claim 1
37[09:37] That day was "one of the worst drops in the whole history of the Korean stock market"factual✅ confirmedNo search of its own — rests on claim 6, whose sources record the 23 June −9.99% triggering a circuit breaker, the first such halt since March 2020, and place it among the KOSPI's handful of worst sessions
38[09:37] "a 10% one-day drop in the US stock market, that would be the third worst day in American history, only behind 1987 and the Great Depression"factual🟠 misleadingThe exclusion list is wrong. 16 March 2020 was −12.0% on the S&P 500 (−12.9% on the Dow) — worse than 10%, and neither 1987 nor the Depression. A −10% day lands 4th–5th on either index, not 3rd
39[13:52] Black Monday [October 1987] was "the worst day in stock market history"factual✅ confirmedS&P 500 −20.5%, Dow −22.6% on 19 October 1987, still the worst single-day percentage loss
40[11:42] Leveraged ETFs "have to rebalance every day, which means at the end of a big down day, the fund has to sell into a falling market to maintain its leverage ratio"factual✅ confirmedLeveraged ETFs reset exposure daily via swaps and futures: "If the underlying index rises, the fund must buy additional exposure at the close; if it falls, it must sell" — structurally short gamma, buying strength and selling weakness
41[11:42] Over two days "more than 600 billion dollars got erased from Asian markets"factual🟡 plausibleNo published figure found for the two-day Asian total. Sanity check from the price data: a −9.99% KOSPI day on a ~$2.1tn market is ~$210bn, a −3.55% Nikkei day on ~$7tn is ~$250bn, plus Taiwan and Hong Kong — so a $600bn two-day figure is in range. Index moves
42[11:42] In the same two days, "Japan went down more than 4%"factual🟠 misleadingNot in those two days. The Nikkei fell −3.55% on 23 June, then −0.88% and +4.61% across the following two sessions. It did clear −4% later in the episode (−4.15% on 26 June, −4.03% on 17 July), so the number exists — attached to the wrong days. No search of its own — computed from the same exchange price series as claim 1
43[12:45] The fall was halted at one point only because "the market was literally closed for a national holiday"factual✅ confirmedKRX closed Friday 17 July 2026 for the newly reinstated Constitution Day. Context fits: the KOSPI fell 6.37% on 16 July and 4.46% on the next open, 20 July
44[14:56] "the official margin data only tracks one kind of borrowing, which is traditional margin loans at brokerage accounts" and misses leveraged ETFs, options, portfolio margin and private creditfactual🟡 plausibleRight in substance — embedded ETF leverage, options notional and non-broker credit are nowhere in the series. One correction: portfolio-margin debit balances are reported to FINRA under Rule 4210(g)/4521, so that item does not belong on the list
45[14:56] Leveraged ETFs hold "hundreds of billions of dollars in them, including 2x funds on single stocks like Nvidia and Tesla"factual⚪ not checked
46[14:56] "We have zero-day options where people gamble on what the stock market does in the next 6 hours. They're all trading at record volumes."factual✅ confirmedSPX 0DTE hit a record monthly-average 3.0 million contracts a day in February 2026, 63% of all SPX options volume, up from ~5% in 2016
47[17:06] Of OpenAI and Anthropic: "they still lose tens of billions of" dollars a yearfactual⚪ not checked
48[20:18] "In the year 2000, Coca-Cola needed 10,000 routers from Cisco. By 2001, they needed only 2,000."factual❓ unverifiable (searched)Two angles — the claim as stated, and the John Chambers customer-cancellation anecdotes from Cisco's 2001 collapse — returned nothing on Coca-Cola, 10,000 routers, or 2,000. Searches were working: they surfaced Cisco's $2.25bn inventory write-off and the real mechanism, where "half of the apparent demand turned out to be multiple orders placed to guarantee delivery". A named company with exact figures and no footprint is the fabrication-tell pattern; it is not proof of invention
49[20:18] "Cisco's earnings collapsed and the NASDAQ went down 78%"factual✅ confirmedNasdaq Composite 5,048.62 (10 Mar 2000) → 1,114 (Oct 2002) = −78%. The "because of that" linking one customer's order cut to the index is his own inference, not a sourced causal chain
50[17:06] "Jim Chanos, the guy who called Enron"factual✅ confirmedReferred to as the Enron short seller in current coverage
51[21:22] "Shrek 5 scheduled to be released June 2027"factual✅ confirmedMoved from 23 December 2026 to 30 June 2027
52[23:30] The vice president of the Titanic's owner said "We have absolute confidence in the Titanic. We believe the boat is unsinkable." after it had hit the icebergfactual✅ confirmed3 URLs → 2 origins: Philip A. S. Franklin, VP of International Mercantile Marine (owner of White Star Line), on 15 April 1912 — "We place absolute confidence in the Titanic. We believe that the boat is unsinkable." If anything he undersells it: the ship had already sunk, hours earlier
53[22:27] "Some analysts genuinely believe that the AI AI has another 6 to 12 months of growth"factual❓ unverifiable (by construction)No analyst named, no firm, no report — nothing to check against
54[21:22] Chanos expects the reckoning "sometime in late 2026 or 2027"predictionDated and attributed; not rateable now
55[07:30] "there are hundreds of companies called data brokers that are legally collecting our personal information" — sponsor segment for DeleteMefactual✅ confirmedDeleteMe alone covers 180+ data brokers, with higher tiers reaching 700–850+ sites. Note the source tier: this is the advertiser's own category framing, corroborated by third-party review sites (tier 3)
56[08:32] "They were just named the number one data removal service by Wirecutter" — sponsor claim for DeleteMefactual❓ unverifiable (searched)Two searches found no Wirecutter ranking of any kind; a third, domain-restricted to nytimes.com, was refused outright — Wirecutter's publisher is not reachable by this search agent, so this is unreached rather than absent. What is visible: 2026 third-party roundups more often place Incogni first
57[08:32] "real people are actually doing the removal work. It's not some outsourced third party" — sponsor claim for DeleteMefactual✅ confirmedDeleteMe "sends real people to manually process your opt-out requests" and uses privacy experts rather than automated scripts. Tier 3 review site; the advertiser is the subject, so treat as corroborated marketing rather than audited
58[08:32] DeleteMe has "removed over 68 instances of someone selling my info" and saved him an estimated 14.5 hours — sponsor segmentanecdote❓ unverifiable (by construction)A private account dashboard belonging to the person paid to describe it
59[05:20] A documentary investor whose "portfolio was worth over $650,000, and she was up 150%"anecdote❓ unverifiable (by construction)Unnamed subject, unnamed documentary, private brokerage account
60[04:16] "one of their biggest finance influencers actually said that their last chance to build wealth is the stock market"factual❓ unverifiable (by construction)No name given, so there is no referent to search

Tally: 60 claims extracted, 44 individually source-checked — 35 confirmed, 5 plausible, 7 misleading, 3 false. 6 unverifiable; 2 not checked; 2 not rateable.

Ambiguous: 2 claims dropped before verification — that hyperscaler spending "happens to be the revenue of Nvidia, of Micron, and of the whole AI trade" (a sum of three revenues, each of them separately, or a notional sector aggregate — the readings settle differently), and "they're saying that the first hyperscaler to cut back on AI spending will be rewarded by the market", where they is never identified.

Unreachable: 4 sources — all blocked (namu.wiki's July 2026 crash chronology, Real Investment Advice's margin-debt critique, Stooq's KOSPI CSV, and Wirecutter/NYT). Named in rows 6, 14, 56; the Stooq block was worked around with exchange price data from Yahoo Finance and cost nothing.

Hype signals observed

To his credit, and it is unusual enough to name: he explicitly retracts the Shrek gag — [22:27] "Obviously, that's a joke and it's my way of saying no one knows what's going to happen and the dates don't perfectly line up" — presents two symmetrical scenarios rather than one, discloses the sponsorship plainly, and says [23:30] "this is not financial advice and I'm not telling anyone to not invest their money". No prompt-injection attempts, no hidden text, no neutralised fences in the transcript.

Incentive analysis

Andrei Jikh is a personal-finance YouTuber with 3.33 million subscribers; this video drew 2.7 million views in under two weeks. Two revenue paths run through it: the DeleteMe sponsorship with an affiliate discount link, and a premium membership sold in the outro. Crash content converts on both — it is high-CTR, and it makes a paid "here's what I'm actually doing with my money" tier feel necessary rather than optional.

The structural problem is that the video is unfalsifiable in his favour. Scenario one is "Korea stays contained"; scenario two is "Korea was just a warning." Between them they exhaust the outcome space, so no future state of the world can make the video wrong — while the framing, the music and the Titanic close all point at scenario two. He is paid for attention either way and carries no position the viewer can check.

Worth weighing against that: he did no promotional trading of his own here, named a specific falsifiable trigger (hyperscaler capex guidance), and the Korea reporting is more careful than most coverage of it.

Bottom line

The factual spine of this video is unusually solid. Nearly every Korea number lands — the 25% fall in 21 days is exact to the day and the decimal, the 1.2 million margin calls, the 3-trillion-won liquidation, the 320,000 wiped-out accounts, the 2%-to-10% liquidation rate, the holiday closure that interrupted the slide, the president who spent a year talking the market up and then convened emergency meetings. The US macro figures hold too: $376bn and $725bn of hyperscaler capex, 4.5% margin-debt-to-GDP at a genuine record, 36% top-ten concentration, 0DTE at record volume, Nasdaq −78%, and a Titanic quote he gets almost verbatim.

What fails is the connective tissue. The single load-bearing error is that OpenAI and Anthropic "make up 70 to 80% of all the AI compute demand" — the measured figure for those two plus xAI is about 21%, and the entire "unprofitable AI companies hold up the pyramid" argument rests on the inflated version. The second is a reversal: he says the US popped Korea's bubble through nerves about Micron's earnings, but Micron fell 13% that day because Korea's own regulator warned about Korean leveraged ETFs, and Micron's earnings the next day were well received. The bubble was popped from inside Korea, which is a better story than the one he tells and a worse one for his thesis. Add a cherry-picked margin-debt chart, a Micron superlative that is simply wrong, and two opening return figures rebased from twelve months to "this year" to make them 2.5× larger.

One thing the report cannot show: the video's central warning was vindicated within days of publication, and then immediately complicated. The KOSPI fell another 10.84% on 28 July — the real, named Black Tuesday, worse than the one he describes — and then rose 17.91% on 31 July, its largest single-day gain in history. A viewer who acted on the video on 20 July would have been right for eight days and then run over.

If you take one thing from it, take the falsifiable part: hyperscaler capex guidance is a real number, published quarterly, and it is the correct thing to watch. Ignore the analogies.

What a hostile reader would hit first

  1. "70 to 80% of all the AI compute demand" (row 16). One search settles it and the answer is off by 3.5×. It is the claim the pyramid diagram is built on, so it does the most damage per second spent, and it is trivially fixable — Epoch AI's number is public and would have supported a weaker version of the same point.
  2. "Micron… is America's biggest chip maker" (row 34). Nvidia is worth roughly 25 times more and is the largest US chipmaker by both revenue and market cap. This one lands hardest because it is the kind of error that makes a viewer re-read everything else. Saying "America's biggest memory maker" would have been true and would have served the argument better.
  3. The causal reversal on Micron (row 19). A reader who checks 23 June finds Korea's FSS warning triggering the US selloff, not the reverse. The whole "we did this to them" section inverts.
  4. "One in every 30 people" (row 26). 1.2 million out of 51.7 million is one in 43. The video does its own arithmetic on screen and gets it wrong by 40%, in the direction that makes the number scarier.
  5. The Coca-Cola/Cisco router story (row 48). Two search angles, nothing. A named company and two exact figures with no trace is the detail a hostile reader will demand a source for, and there does not appear to be one.

run: 20m16s, searches 44, tools 70, coverage 0, per claim 28s