Why Korea’s Market Rout, Higher U.S. Rates, and the Falling Price of AI Belong in the Same Conversation

All year the argument here has been the same. The risk in this cycle is not one theatrical crash. It is a rolling repricing: money costs real money again, and every investment eventually has to pay you back.

January was the structure warning. April was the borrower warning. June was the liquidity warning. July brought the same story to public markets and to artificial intelligence. Three events that were reported separately—a Korean equity rout, a jump in long-term U.S. interest rates, and a fight over the price of AI models—describe one underlying condition.

That condition is simple to state and difficult to live with. When the risk-free return rises, every risky payoff has to justify itself against it, and assumptions that were never tested during cheap money get tested all at once.

One definition first. A circuit breaker is an automatic timeout: when a market falls too far too fast, the exchange halts trading so everyone can catch a breath. Historically, it is rare.

On July 28, Korea’s KOSPI fell 10.8% and tripped a circuit breaker. Samsung fell 13.4%; SK Hynix fell 14.7%. Together they are more than half the index. The next day the KOSPI fell another 6% and tripped another breaker—the first back-to-back activation for the KOSPI, and the third back-to-back activation for the KOSDAQ, affecting both main Korean markets. By July 30, the KOSPI was down roughly 34% for the month, past the prior records set in October 1997 and October 2008.

Even after all that, the KOSPI was still up roughly one-third for the year. At its June peak, it had more than doubled from year-end. That does not make a 34% monthly fall minor. It makes it a violent repricing after a parabolic run. The more useful question is why the market had become so fragile.

Samsung and SK Hynix make memory chips needed by AI systems. Their prices rested partly on two assumptions: that memory stays scarce, and that China’s competitive gap lasts. Then came a report that a Chinese state-owned company had begun producing advanced chipmaking machines—perhaps five this year and twenty in 2027. Those machines still require testing and remain well behind ASML’s Dutch equipment.

The report was a spark, not the whole fire. Lofty valuations, leveraged funds, foreign selling, AI-financing worries, and—the next day—SK Hynix earnings that fell short of enormous expectations all amplified the selloff. Five unproven machines challenged an assumption in a market already primed to unwind.

Same week, different continent. On July 29, the yield on the 30-year U.S. Treasury briefly hit 5.24%, the highest since 2007. That happened even after June inflation cooled to 3.5%, down from 4.2% in May.

One good inflation report did not settle the argument. June’s improvement partly reflected cheaper gasoline, while renewed fighting with Iran threatened to reverse it. The Treasury still has to finance enormous deficits, a constraint the Congressional Budget Office has documented in its long-term projections. And on July 29, the Fed held its policy rate at 3.5% to 3.75% while three policymakers dissented in favor of a quarter-point increase. Markets were left unsure how Kevin Warsh would respond.

Why care if you never buy a bond? The 30-year Treasury does not set every consumer rate, but long-term yields influence mortgages, corporate borrowing, and what investors demand from risky assets. They set the price of patience.

Here is the idea in one sentence: when U.S. government bonds offer about 5%, a risky investment has to promise substantially more—or the risk is not worth taking. A 30-year bond is not a savings account; its price can fall if you sell early. But for an investor who can hold it, a roughly 5% promised return from the U.S. government makes a distant, uncertain payoff less appealing. The market is recalculating.

Current guidance from Amazon, Google, Meta, and Microsoft adds up to roughly $730 billion of capital spending in 2026—almost 80% more than last year’s $410 billion. Much, but not all, is tied to AI and data centers. In fairness, these are strong companies with cash and customers. That is the best argument that the buildout is rational. Now do the math anyway.

A clean break-even estimate is impossible. Buildings last decades; servers and chips have shorter lives; and the $730 billion includes non-AI spending. But the arithmetic survives the imprecision. Hardware must be replaced, data centers must be powered, and hundreds of billions of dollars must earn more than government bonds. The technology can work brilliantly and the investment can still disappoint.

The squeeze is two-sided: required returns are rising while the price of capable AI is falling. Big Tech can fund much of this from cash flow, but cash still has an opportunity cost.

Our firm spends its working life on the other side of this math. We restructure companies that have run out of cash, time, or lender patience, and the recurring pattern is not usually a bad idea. It is a promising business financed on a timetable it could not meet. The same discipline applies whether the conversation is strategic advisory before a covenant breach, a lender negotiation, or a Chapter 11 filing after optionality is gone.

On July 24, twenty-five companies—including Nvidia, Microsoft, and Meta—urged Washington not to restrict “open-weight” AI models. By July 30, more than 230 organizations had signed. Users can download, inspect, modify, and run open weights on their own infrastructure. That can eliminate a subscription, but it is not literally free: hardware, cloud capacity, electricity, and licenses still matter.

Notice who did not sign at launch: three leading closed-model developers—OpenAI, Google, and Anthropic. OpenAI and Google joined later. Anthropic remained absent as of July 30.

This is both a safety debate and a business debate. Closed-model providers emphasize control and risk. Chip and cloud companies benefit from an ecosystem consuming compute. Startups benefit when model costs fall. Nobody has to be dishonest for incentives to shape the arguments.

If access to good-enough AI gets cheap, then parts of knowledge work get cheap. Not gone. Cheap. Drafting a contract. Building a financial model. Writing a first-draft marketing plan. Summarizing a hundred pages of documents. Coding a working prototype. Those tasks helped junior professionals prove their value; a downloadable model can now do a meaningful share of them at three in the morning for the cost of compute.

For companies spending roughly $730 billion, premium prices are harder to sustain when competitors drive model prices toward zero. The buildout faces a rising financial hurdle while its product faces falling prices. That is the squeeze. For individuals, the value is moving away from producing the work and toward judging the work: knowing which answer is wrong, knowing what the client actually needs, and being accountable when it matters. Judgment, relationships, and accountability do not download.

Most people collapse these into one question. Keep them separate and the decisions get better.

• Is the technology real? Yes. That is settled, and it arrived faster than most expected.
• Is there demand? Yes. Companies are adopting it and paying for it.
• Will the company earn a good return? A different question entirely. It depends on what the equipment cost, how long it lasts, and what the money cost.
• Is that return already in the price? The only one that decides whether an investor makes money.

AI can be a yes on the first three and a painful no on the fourth. Korea just demonstrated how that happens—in two days.

The other side deserves a hearing. On July 29, Microsoft reported faster-than-expected Azure growth. On July 30, Amazon reported 37% growth at AWS and raised its spending plan. Both stocks jumped. That is real evidence that demand may be catching up with capacity. If chip demand is structurally higher, July may prove to have been a buying opportunity. Rates could also fall if the oil shock fades.

Building ahead of demand has been right before. We still use fiber networks laid during the dot-com boom. But much of the capital that financed them was destroyed. Being right about the technology and wrong about the price is the most expensive way to be right.

That distinction is the same one at the center of productive borrowing and deferred decisions, and it is the reason private-credit liquidity stress and public-market concentration are two readings of one instrument.

Markets test character because they force a distinction between patience and postponement. Patience is holding to a sound decision when the crowd, the quarter, or the easy money pressures you to abandon it. Intellectual honesty is recognizing when the facts have changed. Character is having the courage to make the hard decision while it is still yours to make.

In almost every failing company we have worked on, there was a moment—usually months before anything visibly broke—when someone in the room understood that the plan no longer worked. Often nobody was lying. The numbers remained defensible, the story remained plausible, and one more quarter might still have saved them. But somewhere along the way, patience became delay and optimism became avoidance. By the time the decision was unavoidable, it was no longer theirs. The lenders, the cash balance, or the market made it for them.

Interest rates are arithmetic. Character is confronting that arithmetic early enough to preserve your choices. That is also the standard we hold ourselves to on The Puck: not simply “what is your thesis?” but “what is the best argument that you are wrong, what evidence would change your mind, and if the facts did change, would you act?”

Last month’s conclusion was that the future rewards the country that invests before the bill comes due. This month is the other half: the future punishes the investor who assumes the bill stays cheap. Korea did not tell us AI is fake. It showed that a market can lose roughly one-sixth of its value in two days when one report challenges a core assumption—and concentration, leverage, and impossible expectations do the rest. The technology was real. The valuation depended on something else. It was cheap money.

Boards and management teams weighing what this means for their own capital structure can review distressed-company options or speak with our team.