South Korea Did Not Pop the AI Bubble. It Stress-Tested It.

On 22 June 2026, the KOSPI-South Korea's main stock index-closed at a record 9,114.55.

That same day, SK Hynix overtook Samsung Electronics as the country's most valuable listed company. It was the perfect symbol of the AI boom: a manufacturer of the fast memory chips used in AI computers had replaced a decades-long national champion at the top of the market.

Three weeks later, the KOSPI closed at 6,747.95.

It had lost approximately 26% from its peak.

The obvious headline is that South Korea's AI bubble popped. But that description misses the most important part of the story.

AI demand did not disappear. The underlying companies did not stop making money. Samsung reported extraordinary profit growth. Micron described intense customer demand and limited memory supply. SK Hynix raised $26.5 billion in a US share sale where investors wanted to buy far more shares than were available.

The business behind the boom remained strong.

The way the market had been built broke first.

That distinction matters far beyond Korea. It shows how the global AI trade could unwind even while the technology continues to improve.

A national index became a borrowed bet on two companies

By June, Samsung Electronics and SK Hynix represented more than half of the KOSPI's value. On some days, the two companies and high-risk funds linked to them accounted for most of the market's trading activity.

The main index contained hundreds of companies, but its behavior increasingly depended on one question:

Will global spending on AI infrastructure continue at its current rate?

That dependency was not imaginary. The computer chips that power AI need enormous quantities of fast memory. Samsung and SK Hynix are critical suppliers. The largest cloud companies-including Microsoft, Amazon, Google and Meta-were expected to spend more than $700 billion during 2026, much of it on data centers and AI infrastructure.

But a sound business story became a dangerously concentrated investment.

This is a familiar engineering problem. One part can work reliably while the system around it remains fragile. If too many processes depend on the same service, one disruption can affect everything-even when that service is working exactly as designed.

South Korea made continued AI spending that shared weakness.

Why so many Korean investors joined the trade

South Korea has roughly 14 million retail investors, known locally as “ants.” The name captures their perceived power: individually small, collectively capable of moving the market.

Their participation did not grow in a social vacuum.

Stock ownership increased from about 6 million people in 2019 to more than 14.5 million by the end of 2025. At the same time, housing in Seoul became inaccessible to many younger workers. One reported average price for an 84-square-metre apartment reached KRW 2.14 billion.

When property stops looking attainable and wage growth does not close the gap, the stock market begins to feel less like an investment choice and more like the remaining path to wealth.

The Korean government encouraged broader share ownership as an alternative to the country's attachment to real estate. New rules gave smaller shareholders more protection. The government also tried to eliminate the “Korea discount”-the tendency for Korean companies to be valued below comparable companies abroad. These efforts gave the rally a legitimate policy foundation.

Then spectacular returns added urgency.

This is where the fear of missing out becomes social. The fear is no longer merely “I might miss a profitable trade.” It becomes “everyone else is moving forward, and this may be my last opportunity not to fall behind.”

That does not make ordinary Korean investors irrational. It makes their willingness to take risks understandable.

It also made the boom easier to fuel with borrowed money.

Borrowed money converted belief into forced action

On 27 May, South Korea allowed new funds linked to Samsung and SK Hynix to begin trading. These funds were designed to deliver twice the daily gain-or twice the daily loss-of one company's shares.

Demand arrived immediately.

By 19 June, ordinary Korean investors had bought approximately KRW 8.2 trillion more of these funds than they had sold. The SK Hynix-linked funds held KRW 9.15 trillion, while Samsung-linked funds held KRW 5.22 trillion.

The amount investors had borrowed from brokers to buy shares was also near a record.

Borrowing changes the behavior of a market because it changes who gets to decide when to sell.

An investor using only their own money can look at a 10% decline and choose to wait. A fund promising twice the daily move must adjust its holdings. A broker must protect the money it lent. When the shares backing a loan fall far enough, selling is no longer an opinion.

It becomes a rule.

The feedback loop looked like this:

The IMF had warned about exactly this kind of mechanism in April. It found that funds promising multiplied daily returns were linked to larger price swings mainly when markets were already under stress. Their automatic buying and selling can turn a large move into an extreme one.

South Korea did not invent this failure mode. It merely ran it at unusually high speed.

What happened during the three-week sell-off

The KOSPI did not fall in a straight line. It alternated between collapses and violent rebounds.

On 23 June, one day after the record, the index fell 9.99%. Samsung and SK Hynix each lost more than 12%, triggering a market-wide trading halt.

The next day, the KOSPI rebounded 3.26% as ordinary investors and large Korean investment firms bought the decline. Foreign investors still sold KRW 4.66 trillion more than they bought.

Then Micron reported exceptional results and said it expected demand to remain strong. Korean memory stocks surged. The business case for AI memory appeared intact.

It did not stabilize the market.

On 7 July, Samsung forecast a 19-fold increase in quarterly operating profit and a third consecutive record quarter. Its shares fell 6.9%. SK Hynix fell 6.1%. The KOSPI triggered another trading halt.

That was the most revealing moment of the correction.

When record earnings no longer lift a stock, the market has stopped asking how strong the current business is. It has started asking how much future success is already included in the price-and how many borrowed bets still need to be closed.

By 8 July, the KOSPI was more than 20% below its June record.

The lesson from SK Hynix's US debut

The next chapter is especially relevant to the global AI market.

On 10 July, SK Hynix made its US market debut. It used depositary receipts, which allow US investors to buy an overseas company's shares in dollars on an American exchange. The company raised $26.5 billion. The new receipts opened 14% above their initial price and finished the day almost 13% higher.

Global investors still wanted AI exposure.

In the next Korean trading session, SK Hynix's local shares fell roughly 15%. Samsung lost about 10.7%. The KOSPI dropped 8.95% and trading was halted again.

The listing was a success. The local market still broke.

This is not a contradiction. Strong demand for newly issued shares can exist while existing shareholders are rushing to reduce risk. A new listing may attract long-term investors while others take profits, move money between the US and Korean shares, or sell because borrowed money leaves them no choice.

A blockbuster offering can also become the last piece of good news the trade was waiting for.

In market language, it becomes a “sell the news” event. In engineering language, the deployment succeeds while the surrounding system fails under load.

Regulation arrived after the chain reaction had started

On 16 July, Korean authorities announced restrictions on funds that multiply the daily moves of individual companies.

They temporarily stopped new fund launches and promotions, strengthened investor education, introduced tighter controls when a fund's market price moved too far from the value of its holdings, and tripled the minimum cash required to participate from KRW 10 million to KRW 30 million.

The measures reduced the amount of new fuel entering the system.

They could not instantly remove:

• existing high-risk fund positions;
• money investors had already borrowed;
• funds listed in Hong Kong and other markets;
• automatic daily trades by funds already on the market;
• the KOSPI's dependence on two companies.

That is why the volatility continued.

On 15 July, the KOSPI rose 6.24%. The following day, it fell 6.37%. On 20 July, it dropped another 4.5% to its lowest close since April. On 21 July, it rebounded 3.56%, briefly triggering a safety mechanism that pauses automated trading when it moves too quickly.

This was not normal trading. The market was still processing a huge pile of borrowed investments, producing extreme moves in both directions.

JPMorgan estimated on 21 July that investors had completed approximately 75% of the reduction in these high-risk funds needed to return them to what the bank considers a sustainable size. That suggests much of the automatic selling may already have happened.

It does not prove that the market has found a bottom.

Is this a preview of the global AI market?

Not exactly.

The US market is deeper, more diversified and supported by companies with enormous cash flows. The top AI-related businesses do not represent more than half of the entire market in the way Samsung and SK Hynix came to dominate the KOSPI.

But the chain is similar:

The companies are different. The products are different. The common assumption is the same.

The amount US investors had borrowed from brokers reached a record $1.502 trillion in June. Trading in contracts that bet on the S&P 500's movement over a single day also reached record average volume. Neither number predicts a crash by itself. Both show how quickly a change in expectations can spread.

The critical trigger will not be an announcement that AI has failed.

It will be a change in how markets react to AI spending.

Today, Microsoft, Amazon, Google and Meta are often rewarded for increasing AI infrastructure spending because it signals leadership in the race. I am watching for the first of them to spend less-and see its share price rise because of that decision.

That would mean investors have stopped rewarding expansion and started rewarding restraint.

The same spending cut could leave more cash at the technology company while reducing expected sales for chip designers, memory suppliers, data-center operators and energy providers.

One company's decision to spend less becomes a supplier's lost sale.

My prediction: the IPOs come before the decisive break

Anthropic submitted confidential IPO paperwork on 1 June. OpenAI followed on 8 June.

Neither company has committed to a final listing date. OpenAI has explicitly said that the timing remains undecided.

My prediction is still confident:

The global AI bubble will not decisively burst before public markets get the OpenAI and Anthropic IPOs.

The reason is not that anyone can schedule a market crash. The reason is that public investors have not yet been asked to put a daily price on the two companies that best represent today's AI boom.

These companies are the clearest expression of the belief that increasingly powerful AI models will justify enormous investment. Their public documents will reveal details that private companies can currently discuss selectively: computing bills, reliance on a few major customers, the cost of delivering their services, how quickly they spend cash, and whether they have a believable path to lasting profits.

If public investors absorb both listings at ambitious valuations, the cycle gets one more extension.

If the deals are delayed, offered at lower prices than expected, or fall after their debut, the market will have answered a more important question than any AI performance test:

How much of the promised AI future is it still willing to finance today?

South Korea does not invalidate this prediction. It shows how the final act may look.

SK Hynix completed a historic offering during a violent correction. Its debut succeeded. The surrounding trade still suffered another collapse immediately afterward.

The IPO can be the climax. It is not protection from what follows.

The engineering lesson

The wrong conclusion from Korea is that AI is fake.

The right conclusion is that strong parts do not rescue a fragile system.

Samsung can report record profits. SK Hynix can sell all the memory it can produce and complete a historic share sale. The world's largest technology companies can keep building data centers. Every individual decision can remain rational.

The system can still be fragile if every participant depends on the same assumption, and borrowed money converts a small change in expectations into forced action.

That is true in stock markets.

It is also true when building AI products.

If your product depends on one AI provider, one cloud company and one source of funding continuing to deliver forever, you have not protected the business from failure.

You have hidden one shared weakness behind growth.

South Korea made that dependency visible.

The rest of the world should treat it as a warning, not a prophecy.