Written at a time when faith in technology is tottering

First, a declaration:

1. I hold positions, which were not established recently, and I have followed this round of market for more than three years.

2. I do not go all-in, and I use an allocation method that I can afford to withstand.

3. No leverage — I am relatively conservative, and I do not think I know everything well, always believing that things are not that simple.

4. I reduced my positions in June, but I held on firmly through March and April; the gains in May and June made me feel that holding on was not a problem in early July, so I did not liquidate all positions with one click.

5. I also hold dividend, quantitative and pharmaceutical assets. I never believe that technology is the opposite of all other assets.

So I am not writing from the sidelines. I am not a tech hater, nor am I someone waiting to see a joke. What I write is the remaining judgment after paying my tuition fees in the market.

There are three questions, still the same three: What went wrong? Why? What on earth am I believing in?

First, the conclusion

There are many explanations for this round of decline in October, and the most convenient one is "the AI bubble has burst". If you see it this way, it may be too crude and not rigorous.

TSMC's net profit in the third quarter rose 68.7% year-on-year, hitting a record for the sixth consecutive quarter; Micron's single-quarter revenue more than tripled year-on-year, with a gross margin of 87%; the sum of next year's capital expenditure guidance from the four cloud manufacturers is about 745 billion US dollars, more than 70% higher than this year. If demand is really disappearing, these figures should not look like this.

What went wrong is at the other end: the money in this construction is getting more expensive. And all grand narratives about technology will eventually fall on a very specific action — paying interest.

Faith rarely collapses because of a single piece of bad news. It is that one morning, a debt matures, and the whole structure shakes a little.

Those days

When talking about October, we cannot bypass September. September looked too much like a successful repair, and it was exactly that month that made everyone let their guard down again.

On September 22, the Nasdaq hit an all-time high during intraday trading; Apple closed at a new high, with its market value approaching 5 trillion; Meta rose nearly 13% in a week because of an AI agent called Muse; the Philadelphia Semiconductor Index rose 6.27% in the week ending September 25, posting four consecutive weekly gains, the longest streak since May.

The A-share market rose accordingly. On September 16, the Sci-Tech Innovation 50 Index rose 4.14% in a single day, and the entire semiconductor chain exploded. One of the reasons at that time was the "15th Five-Year Plan" for the Development of the Electronic Information Manufacturing Industry jointly issued by the Ministry of Industry and Information Technology and the National Development and Reform Commission, which stated that the revenue of enterprises above designated size would exceed 30 trillion yuan by 2030. By September 22, the Sci-Tech Innovation 50 Index touched 1707.67 points during intraday trading.

In those days, the "AI bubble theory" was almost unheard of in public opinion. Everyone was calculating how much more it could rise.

Then October came.

Money moved first. On October 7, the 10-year U.S. Treasury yield hit 5.364% during intraday trading, the highest level since 2002; the 30-year yield reached 5.696%. The third quarter was the worst quarter for U.S. Treasuries since 1994. Here is a background supplement: the Federal Reserve really raised interest rates by 25 basis points on September 16, to 3.75%-4.00%, the first time since July 2023, with all 12 votes in favor. Walsh's exact words were "Inflation is too high, and it has lasted too long". In the dot plot, 16 out of 18 people believed that there would be another rate hike within the year, the median forecast for the interest rate at the end of 2026 was raised from 3.8% to 4.1%, and the time point when inflation returns to 2% was pushed to 2029. Trump shouted that day that interest rates "should be cut to 1% or lower". The market got the message, and the 10-year U.S. Treasury yield later stood back above 5%.

Then came inflation. The September CPI announced on October 14 was 3.7% year-on-year, compared with the previous value of 3.4%; the core CPI was 2.6%, month-on-month 0.6%, and gasoline alone rose 9.1% month-on-month, accounting for more than half of the month's increase. The market's pricing of a rate hike at the September 28 interest rate meeting jumped from 22.1% to 43.7% in one day. The next day, the September PPI came out at 6.3% year-on-year, compared with the expected 5.3%. Later, employment data weakened: 29,000 new non-farm jobs were added in September, compared with the expected 90,000, and the unemployment rate rose to 4.2%. As a result, the probability of a rate hike in October fell back to around 18%, but the pricing of a rate hike in December was still at 70%. To put it bluntly, they are in no hurry to raise rates, but they also have no intention of stopping.

Oil prices are still climbing. The Strait of Hormuz has not returned to normal. On October 11, 6 ships were attacked in one day, and the oil transportation volume through the strait in the past week was about 43% less than before the incident. Brent crude oil is between 100 and 107 US dollars, and WTI is around 95 US dollars. The president of Saudi Aramco said that global commercial crude oil inventories have fallen below 6 billion barrels, which is at an "extremely dangerous" low level.

Then came the memory sector. According to TrendForce, the conventional DRAM contract price in the fourth quarter will rise by another 10%-15% month-on-month, the overall DRAM including HBM will rise by 15%-20%, and NAND will also rise by 15%-20%; the average mixed price of HBM in 2027 is expected to rise by 121% year-on-year. Micron's revenue in the fourth quarter of FY2026 was 54.229 billion US dollars, up 31% quarter-on-quarter and 379% year-on-year, with a Non-GAAP gross margin of 87%. Data center SSD revenue was about 10 billion US dollars, more than ten times that of the same period last year, and the company has locked in more than 75% of its production capacity in 2027.

The same thing looks like a money printer for the upstream, but a life-threatening burden for the downstream. This is probably the most easily misread point in this period.

Then came the crash.

On October 8, the first trading day after the A-share market holiday, the Sci-Tech Innovation 50 Index fell 4.82%, the Wind Optical Chip Index fell by more than 8%, Source Photonics, Changguang Huaxin hit the 20cm limit down, and Sijia Photonics fell by more than 17%. The trigger was a research report during the holiday: Morgan Stanley judged that the FCC might start to restrict Chinese-made optical modules in phases from the 3.2T generation, and one of the paths was to implement "U.S. content exemption", which means that the value proportion of U.S. enterprises in the BOM must reach 65% to be approved. Coupled with the rumor of "optical chip price cuts", the entire optical communication chain was suppressed.

From October 13 to 15, it was the turn of the U.S. stock market. The Philadelphia Semiconductor Index retreated more than 20% from its high point since June, officially entering a technical bear market again, erasing more than 2 trillion US dollars in market value in seven trading days. On the 15th, the SOX index fell 4.01% in a single day, AMD and Intel fell by more than 7%, SanDisk fell 8%, Marvell and Western Digital fell by more than 7%, Micron and SK Hynix fell by more than 5%.

On the A-share side, the Sci-Tech Innovation 50 Index fell from 1530.01 points on September 30 to 1377.63 points on October 16, down 9.96% in the range, and the semiconductor index fell 10.69% in the same period.

On October 16, the Dow rose 556 points, but the Philadelphia Semiconductor Index still closed down 0.88%, falling for the seventh consecutive trading day. Money flowed out of semiconductors and went to healthcare, consumer and dividend assets.

The lesson from South Korea

When talking about the most extreme sample of "faith teetering on the edge", it must be South Korea.

In July, the regulator raised the minimum margin for chip leveraged ETFs from 10 million won to 30 million won, and only accepted cash; foreign investors net sold 32.37 billion US dollars in June, and sold another 9 billion to 11 billion US dollars in the first half of July. In less than a month, the two memory giants fell nearly 50% from their highs, and 320,000 to 460,000 leveraged accounts were forced liquidated. That was the first round of clearing.

In September, the market gave them another chance. The KOSPI fell back from 7000 points on the 9th to 6700 points, and retail investors rushed in again: in the week from September 10 to 17, they net bought 256.7 billion won of "KODEX Leverage" ETF, 1.6 trillion won of Samsung Electronics, and 1.3 trillion won of SK Hynix. They bought spot assets, as well as double-leveraged single-stock ETFs, betting on a rebound.

In October, the bill came. On the 15th, the KOSPI opened at 5673.01 points, down 5.59%, and fell nearly 8% during intraday trading; on that day, 14 single-stock leveraged products of Samsung Electronics and SK Hynix in South Korea fell by 19% to 22% in a single day. Based on the issue price, the KODEX SK Hynix single-stock leveraged product fell from 23,450 won to 5,285 won, down 77.46%; similar TIGER products fell 77.63%. If you invested 10 million won at the time of listing, the current valuation is about 2.25 million won. These two batches of products have only been listed for five months.

By October 13, foreign investors had accumulated a net sale of 203.6977 trillion won on South Korea's main board this year, equivalent to about 141.5 billion US dollars, of which Samsung Electronics and SK Hynix together accounted for about 176 trillion won, accounting for 86%; the foreign shareholding ratio of Samsung Electronics fell to a new low since 2008.

As if there are seven hands

I always believe that it is too hasty to explain the market with conspiracy theories and a single force. When the AI narrative becomes a global high-stakes competition that affects the whole body by moving one part, it cannot be a one-dimensional matter.

There are at least seven hands hanging over it, or six mechanisms moving at the same time.

The first hand is the power of short selling and arbitrage.

This time, its cutting edge has shifted from valuation to credit. On October 13, the yield of a CoreWeave bond maturing in 2031 once rushed to 16%; the 5-year CDS spread hit 1194.5 basis points during intraday trading, implying a default probability of about 60% in the next five years, compared with 30% a year ago. Musk only replied two words: "yikes". On the same day, Oracle's 5-year CDS was at 288 basis points, SoftBank's expanded to 424 basis points, and the CDS of Amazon, Google, and Meta all reached historical highs. On October 12, BlackRock downgraded its long-term "overweight" U.S. stocks to "neutral", while abandoning its 6-year main tone of "underweight U.S. Treasuries", on the grounds that the fluctuation risk of the AI construction cycle is greater than the opportunities.

The second hand is liquidity.

The tap has been almost completely turned off. The 10-year Treasury note at 5.364% and the 30-year at 5.696% are both 20+ year highs, and the U.S. dollar index has risen all the way from 100.3 to 102.7. People thought they were betting on the prospects of technology, but in fact they have been betting on the cost of capital.

The third hand is the tech giants. Amazon raised its 2026 capital expenditure from 200 billion US dollars to 220 billion, Google raised it from 180-190 billion to 195-205 billion, Meta raised its lower limit from 125 billion to 130 billion, Microsoft is about 175-190 billion, and TSMC's 2026 capital expenditure is 60-64 billion. The sum of the four cloud manufacturers is about 745 billion, up about 75% year-on-year, and Wall Street expects it to exceed 1 trillion in 2027.

But the scary thing is: these super giants can no longer be self-sufficient even with growing performance.

Goldman Sachs expects that hyper-scale cloud manufacturers will issue 420 billion US dollars in bonds next year, a record high; JPMorgan estimates that the five major cloud manufacturers plus NVIDIA will issue about 320 billion US dollars in bonds this year, of which the long-dated part is equivalent to about 303 billion in 10-year equivalents, equivalent to 68% of the U.S. Treasury's new long-dated borrowings in the same period. There are more off-balance sheet items: according to Morgan Stanley's statistics, the off-balance sheet commitments of the five cloud manufacturers plus NVIDIA and Broadcom total more than 3.1 trillion; NVIDIA's equity investment in customers has reached about 99 billion US dollars, compared with about 40 billion at the beginning of the year and about 7 billion a year ago; Broadcom has set a loan limit of 420 billion for Anthropic, for the purpose of buying Broadcom's own chips. So this hand is not tightening orders, it just moves the risk from the customer's procurement budget to the customer's balance sheet.

The fourth hand is geopolitical and political issues.

The Strait of Hormuz is still unobstructed, and Brent crude oil has returned to 100-107 US dollars. There is a new front on the chip side: in early October, bipartisan U.S. senators introduced a bill to ban federal sensitive systems from purchasing Chinese-made optical modules, and Morgan Stanley further deduced that the FCC would phase in regulation from the 3.2T generation, using "65% U.S. content" as the exemption threshold.

The fifth hand is the power of open source.

The actions on the open source side are more solid: DeepSeek released V4.1 Flash on September 10, with a total parameter of 552B, only activating about 8B per token, and the input price is 1 yuan per million tokens during idle hours; on September 30, it open sourced the infrastructure components for Huawei Ascend computing power platform, including compilation tools, computing libraries, and communication libraries, which allowed domestic computing power to move one step from inference adaptation to the training side; Moonshot AI's Kimi K3, with a total parameter of 2.8 trillion, has become the world's largest open source weight model at present, with annualized revenue exceeding 1 billion US dollars.

The sixth hand is human greed and fear.

The premium rate of technology ETFs was once bought to 24.6%; South Korean retail investors poured 256.7 billion won into leveraged ETFs in a week; then the same products fell by 77% from their issue price. From greed to fear, 20 trading days are enough.

The seventh hand is the financing structure itself. This one is the easiest to ignore, and the most fatal. It is not emotion, nor policy, but a balance sheet that must pay interest on schedule