On Thursday the US chip index fell 6.7%. All 30 members closed lower, about $200 billion of value disappeared, and it was the index’s lowest close since late July.
The strange part: almost every piece of demand news that day was good.
The demand side said “more”
- Lam Research (LRCX), which makes chip-making machines, had just said 2027 equipment spending should pass $200 billion, and that parts shortages held its own outlook below demand.
- Western Digital (WDC) beat on revenue, guided next quarter about 12% above expectations, and said its price per terabyte rose 28% in a year.
- NVIDIA (NVDA) CEO Jensen Huang said demand from the new AI cloud providers is “off the charts”, and rejected forecasts of a memory glut by 2029.
Every supplier said the same thing: customers want more than we can ship.
The money side said “wait”
In the same 24 hours, three different players described a different constraint:
- Meta Platforms (META) spent more on investment last quarter than its operations generated, and created a new non-voting share class it can sell to raise money. The stock fell about 7% on Thursday.
- Huang himself said the missing piece for fast-growing AI clouds is working capital, not customers.
- SK hynix cancelled the planned US listing of its storage unit, citing changes in financial market conditions.
The bond market agreed. The US 10-year yield fell about 0.17 points in one day, its biggest drop since 2023. That is what a growth and credit scare looks like, not an inflation scare.
My reading: the bottleneck is moving from wafers to the balance sheets that pay for them.
What this changes (and what it doesn’t)
It changes the ranking, not the direction.
First in line: suppliers paid from customers’ operating cash flow. Chip foundries like Taiwan Semiconductor (TSM) and memory makers with long-term contracts signed with cash-rich cloud giants.
Last in line: builders that fund AI with borrowed money. That is why my exit rule on Oracle (ORCL) is a weekly close below about $130. It was around $126 before Friday’s open.
Two of the week’s best reports (Lam and Western Digital) were sold anyway. When sold-out companies fall on good news, that is usually positioning being unwound, not orders being cancelled.
What changed in the portfolio
- Twelve positions in copper, rare earths, lithium, China, a chip-equipment ETF, Corning and GE Vernova were reset: closed and reopened at full size. That books the loss, averages into the dip and keeps every line large enough to copy. No change in themes.
- Nothing in memory was touched. The rule is still to wait for SK hynix (27 Oct) and Samsung (29 Oct).
- Hedges did their job: the Treasury position (IEF) and utilities, staples and health care rose. Gold did not; it fell about 2.8% to its lowest since July.
- The portfolio fell about 2.2% from Thursday morning to Friday morning.
What I watch next
- Friday’s weekly closes: Oracle versus $130, TSMC versus $430 (no add this week), Micron (MU) versus $858. If Micron closes the week below $858, I still do nothing before the 27/29 Oct reports.
- 27 Oct SK hynix, 28 Oct Fed decision plus Microsoft and Alphabet, 29 Oct Amazon and Samsung. The question for the cloud giants: do they pay for next year’s spending from cash flow, or from new debt and shares?
- Still a maximum of one new AI purchase per week until after the Fed.
Risks, openly
- If the cloud giants also talk about funding rather than demand, the selling can spread to the strongest names.
- Memory and Korea are about 10% of the portfolio, with five major reports inside 72 hours.
- Oil is at five-month highs on Gulf attacks; gold is not acting as a safe haven this month.
The takeaway
Loud days are the ones the rules were written for. Thursday tested four of them at once: who pays the invoice, the bond hedge, the dated memory test and the weekly-close discipline. None needed rewriting. I would rather wait four sessions for real answers than trade a guess two days early.
Portfolio snapshot: Friday 23 October 2026, 10:10 UTC (US prices = Thursday close plus Friday pre-market).
Sources: Tech Times on the chip selloff, Yahoo Finance on Huang’s remarks, Western Digital results coverage, Seoul Economic Daily on Solidigm.
This is not financial advice. Past performance does not guarantee future results, and copy trading carries a real risk of losing money. Always do your own research.