The Shape of a Bubble Shapes Its Deflation

First, look at the day's pattern. On the same day as Meta's earnings, Lam Research issued next-quarter revenue guidance of USD 8.6 billion ± USD 0.4 billion. The next day, Lam Research fell −11.76% and Meta −6.98%. As AI hardware and growth stocks fell, value stocks, consumer staples and Treasuries rose.

Relative performance makes it an even rarer day. Value (IVE) outperformed growth (IVW) by 3.28 percentage points. The last larger gap was Jan 27, 2025 (+4.60 percentage points), when news of China's low-cost DeepSeek AI model sent AI semiconductor stocks sharply lower. The small-cap ETF (IWM, +0.07%) outperformed the Nasdaq 100 ETF (QQQ, −2.86%) by 2.93 percentage points.

Korea reacted before the US market did. Meta and Lam Research reported after the US close on 10/21; Korea's 10/22 regular session opened afterward. The KOSPI fell −3.52% to 5,533.77, SK hynix −5.49%, and Samsung Electronics −4.03%. The KOSPI is 39.3% below its 6/22 peak of 9,114.55.

Oil rose on the same day. Brent reached around USD 113 a barrel (reported closes differ slightly, at USD 112.91–113.17). Inflation pressure remained even as the AI boom cooled. This leads to the second path in Section 07.

The next day (10/23). The KOSPI closed at 5,550.30 (+0.30%). In the US, S&P500 fell −0.52%, Nasdaq −0.21%, and Dow −1.23%; the semiconductor index rose +0.44%. The 10-year Treasury yield was 5.00%. These figures are not added to the charts.

Reading (interpretation). This was the pattern on a cooling day. Producers (semiconductors and equipment) and heavy spenders fell most. Lower rates supported Treasuries and dividend stocks. Areas that had risen less fell less. One day cannot settle the question, but it resembled the historical pattern examined in Section 02.

We place six booms and their endings side by side. Their shapes, triggers, largest losses, resilient assets and what remained.

Lessons to carry over (interpretation).

• What ended booms was usually more expensive money, rather than failed technology. Commodity and inflation shocks, such as wheat in 1847 and oil in 1973, made money expensive. Disappointment in technology then determined the depth of the decline.
• Who held the risk determined what followed. In shareholders' hands, losses were deep but contained; in banks and debt, they left a 10-year wound.
• Producers fell most; users of the remaining infrastructure benefited later. The next generation of internet companies grew on surplus fiber.
• Concentrated bubbles left the rest less damaged. Nonrailway stocks were much less affected, and small-cap value rose 24% in 2000–02.
• Bonds cushioned losses only when inflation allowed it: +33.5% in 2000–02 and −13% in 2022.

So 1847 and 1973 may be closer parallels than 2000 (interpretation). In 2000–02, the Fed could cut rates by more than 4 percentage points. Today, the Fed raised rates in September, and oil is above USD 100 a barrel.

This boom is narrow and deep. In the US S&P500, the top 10 stocks accounted for about 41% at end-2025 and about 38% in June 2026 (secondary sources). The 2000 peak was 27%. Korea is narrower still. Samsung Electronics and SK hynix rose from about 21% in early 2019 to 57.1% on June 25 this year.

The source of money is changing. Initially, companies built with earned money. Borrowing and new shares are now increasing.

• Hyperscalers (large cloud companies) issued USD 219 billion of investment-grade corporate bonds this year through August 20, including USD 62 billion in other currencies. They are expected to account for about 9% of dollar investment-grade issuance in 2026, versus 2% in 2022–24.
• Their bond spreads widened 30bp; investment grade overall widened 2bp. Off balance sheet, they have about USD 1.1 trillion in leases and USD 1.5 trillion in purchase commitments (J.P. Morgan Asset Management, 8/20).
• Meta's Q3 free cash flow was −USD 2.12 billion. Its 2026 capital expenditure guidance is USD 135–145 billion.
• New shares are increasing too. Alphabet sold USD 45 billion of shares in June and said it planned to sell another USD 40 billion. Meta is creating nonvoting shares (effective January 2027). A draft of Anthropic's listing-preparation documents was also reported.

The market has become faster too. In July, the share of same-day-expiry options in S&P500 options trading was 66.2%. In June, the pattern-day-trader rule limiting frequent trades by small investors was removed. On October 29, 1929, New York's ticker tape was 152 minutes late. Today, news and orders move almost together. On October 15, the KOSPI fell 7.6% in a day and circuit breakers were triggered in both markets.

Reading (interpretation). The builders are neither the railway promoters of 1845 nor the telecom startups of 2000. They are established leaders with substantial cash and high credit ratings. But some companies are moving into negative free cash flow, while risk spreads from equities to corporate bonds, private credit and leases. That makes it less visible. Information asymmetry is increasing.

The US 10-year Treasury yield rose from 3.97% on February 27 to 5.31% on October 5. That was the highest since May 2002. Over the same period, the term premium rose from 0.46% to 1.10%, its highest since April 2010. That is about half the rise of 1.34 percentage points in the 10-year yield. The term premium is the extra compensation investors demand to hold long-term bonds (a model estimate).

Central bankers say this directly. Fed Chair Warsh said AI investment hits demand before supply (translated). ECB policymaker Kazaks described a global fight for money (translated), calling AI investment one of the main forces raising its price. The Fed and ECB both raised rates in September (Reuters 10/16; quotations translated).

There is evidence in the opposite direction too. PIMCO counted 43 days with major AI model releases since 2023. The 5-year, 5-year-forward real yield fell a cumulative roughly 1 percentage point on those days; half came from the term premium. AI investment increases demand for money and pushes rates up. AI capability can lower rates by changing expectations for growth and inflation (interpretation).

Inflation and oil push the rest. US September consumer prices rose 3.72% from 1 year earlier (August: 3.35%). The effective federal funds rate rose from 3.63% to 3.88% on September 17.

Reading (interpretation). About half the rise in the 10-year yield was the term premium. That premium is sensitive to bond supply and who buys it. It may therefore have moved alongside AI funding demand growing through corporate bonds, leases and private credit. But these figures do not isolate AI's share. The other half is policy and inflation. If the AI boom cools, the term premium could fall quickly. Oil and inflation determine the policy-and-inflation component. We will revisit the source of the −0.17 percentage points on 10/22 after term-premium data are published, about 1 week later.

Do rates fall when AI shakes? We compare tech's relative return (Nasdaq daily return − Dow daily return) with the daily change in the 10-year yield. A positive 60-trading-day correlation means yields fell on days tech declined. A negative correlation means tech fell when yields rose.

The pattern is clearer when we isolate large tech declines. We define event days as days when tech's relative return is in the bottom 10% of the previous 250 trading days (an observation rule). There were 24 event days this year.

Similar stock declines have come with opposite yield moves. On June 5, tech's relative return was −2.83 percentage points and yields rose 0.08 percentage points. On October 22, the figures were −2.62 percentage points and −0.17 percentage points. On the DeepSeek shock day, Jan 27, 2025, they were −3.72 percentage points and −0.10 percentage points.

Reading (interpretation). In the first half, tech often fell when yields rose: rates appeared to be in the driver's seat. In October, yields fell when tech shook: AI news appeared to be driving. The observation that yields fell when AI risk was discussed fits October's pattern. But there were only 4 October event days, so it is not conclusive. The driver can change again. We record this daily with the first needle in Section 08.

Neglected assets have been turning up from an October 7 low. These are ratios of equal-weight S&P to cap-weight, value to growth, and small caps to Nasdaq 100. All three had declined since late March and reached their lows on October 7, two days after the 10-year yield peaked.

Could this be an opening for small technology companies? History offers a strong precedent. After the telecom bubble burst in the early 2000s, the cost of renting installed fiber collapsed, and the next generation of internet companies grew on it. But three conditions are needed: cheaper capital, cheaper computing, and their own customers and margins.

Only the first condition arrived slightly on 10/22. GPU rental is more expensive than its low 1 year earlier (USD 2.88, 10/22). The software ETF (IGV), up +2.9% this year, lagged semiconductors (+51.5%) heavily. Even since cooling began on 10/5, it is down −0.9% and has not recovered.

Reading (interpretation). Small companies' opening comes when builders reduce investment and computing becomes abundant. GPUs age far faster than fiber, so any surplus may be shorter-lived. Power, sites and transmission grids last longer. Thin businesses reselling other companies' models may shake in the same cooling.

There is more than one path for the boom to cool. History and today's numbers suggest three paths. We cannot choose one now. But we can record which way each number leans every day.

Reading (interpretation). On 10/22, an A-type day occurred against a B-type background. Korea's weight in producers (memory) makes the initial shock large on any path. But on path A, users of remaining infrastructure and beneficiaries of lower rates could return later. That recovery takes longer on path B. Signals of C are visible only in AI borrowers' bonds so far.

We cannot know an enormous complex system in full. We record several needles each day instead. We keep the needles separate. We do not combine them into a single score.

Dates to check.

What we still do not know.

• Whether hyperscalers will cut or raise 2027 capital expenditure. Meta said only that it would keep investing aggressively (translated).
• The size of risk accumulated in private credit and leases. Public information reveals only part of it.
• Whether oil and inflation will permit rates to fall.
• When GPU rental prices will turn.

Reducing the flame does not make a hot-air balloon fall at once. The wind and cargo determine where it lands. In this boom, the wind is rates and inflation; the cargo is narrow concentration and borrowed money. We do not try to predict the landing day. We record the wind's direction and the load's weight each day.