The AI selloff that started in memory chips on Monday had reached data center lenders, power suppliers and the biggest names in the Nasdaq 100 by Friday. The Nasdaq 100 lost −5.49%, its worst week since April 2025. The Dow gained +0.67%, and 350 of the S&P 500’s 503 members finished the week higher.

Every demand signal pointed up. TSMC reported a record quarter and lifted its outlook, and ASML said some customers want to place orders for 2029 and 2030. The pressure came from the cost of money for the companies building AI capacity, and investors spent five sessions repricing it. Inflation ran hot, the Fed signaled it would wait anyway, and on Friday a rumor about a forced seller sent the VIX up by more than almost any similar drop in the past three years.

Weekly Index Performance

Cross-Asset Snapshot

The Week by Theme

Sorting the week’s moves by theme shows where the money left and where it went. Among the Magnificent Seven, only Apple rose, +2.0%, and it closed above $5 trillion in market value for the first time on Wednesday, according to iThinkDiff. Tesla fell −7.3% and Nvidia −6.6%. On Thursday, Apple and Meta said they would team up on AI models that run locally on laptops rather than in data centers, an announcement analysts said added to the worries around data-center debt, The Wall Street Journal reported.

The heaviest losses were in the companies that build and finance AI capacity. CoreWeave fell −25.2%, Nebius −25.1%, IREN −23.3%, Dell −22.9%, its worst week on record, and Bloom Energy −21.1%. Chip equipment and memory followed: Applied Materials −19.3%, Seagate −18.4%, SanDisk −17.9%, Lam Research −17.2% and Micron −14.6%. The Philadelphia Semiconductor Index lost −11.41% and fell eight sessions in a row, its longest losing streak since 2008. The VanEck Semiconductor ETF had its worst week since April 2025.

At the other end, Walmart, Coca-Cola and Eli Lilly each gained more than 6.5%, and Chevron and Exxon Mobil rose about 6%. The winners were defensive companies and oil producers, with no AI story attached.

Funding Fears Drove the AI Selloff

The week’s first warning came from a chip CEO. Speaking at a panel in Singapore on Monday, Broadcom’s Hock Tan said his company is working around wafer and memory shortages and that “capital will be the biggest constraint,” Reuters reported. Coming from the head of one of the main suppliers of AI chips, that was a statement about his customers’ balance sheets.

Two days later the bond market made the same point. On Wednesday, CoreWeave’s 9.625% notes due 2032 yielded about 14.8% and briefly touched 15%, according to investingLive. The stock fell −15.5% that day, and the selling spread to Oracle (−8.8%), Nebius (−10.1%) and Dell (−11.3%). CoreWeave rebounded +5.9% on Thursday after Nvidia’s Jensen Huang told Bloomberg in Stockholm that the company’s cost of capital is falling and that it needs no backstop. Its five-year credit default swaps hit a record 1,194.5 basis points the same day, The Next Web reported, citing ICE data. The stock bounced while the credit market kept pricing more risk. CoreWeave carries at least $35 billion of debt, according to FactSet figures cited by The Wall Street Journal, and one of its bonds due 2031 traded more than 10 percentage points above Treasuries, the threshold at which bonds are usually treated as distressed. By Friday the selling had reached data-center bonds with investment-grade ratings, and CoreWeave’s stock fell another −11.9%.

The demand side looked fine all week. TSMC’s third-quarter revenue rose 41% to $46.7 billion, it guided fourth-quarter sales well above estimates, and full-year growth is now expected around 44%, per Hargreaves Lansdown. ASML reported record quarterly sales of €11.7 billion and said in its results release that its planned supply for 2027 and 2028 won’t fully meet customer demand. Neither stock was spared. TSMC lost −9.0% for the week and ASML −7.8%.

The AI labs themselves added to the debate. Anthropic told prospective IPO investors that third-quarter revenue rose about 45% from the second quarter to $16.8 billion, Reuters reported, via investingLive. The figures are unaudited and come from one investor’s account. They also contain a detail that got less attention. Revenue grew by about $6.9 billion from the first quarter to the second and by about $5.2 billion from the second to the third. The growth rate is still enormous, but the dollar gain was smaller than the quarter before.

Paul Tudor Jones added the supply side in a CNBC interview. He said net equity supply, meaning IPOs, secondary sales and lockup expiries minus buybacks and takeovers, is the most negative since at least 2009, as SpaceX and possibly Anthropic and OpenAI come to market. He expects a fourth-quarter rally after the midterms and said he plans to be short stocks for all of 2027.

The AI selloff looks like a repricing of financing risk. Every company in the chain still reports more demand than it can serve. The weak links are the borrowers in the middle, the neoclouds and data center developers that fund GPUs with high-yield debt and lease them out. When their bonds trade like distressed debt, investors start asking who pays for the next round of capacity, and that question reaches Nvidia, the memory makers and the power suppliers in the same afternoon. Friday’s recap covered the hedge fund liquidation rumor and VARA’s disclosed holdings. The rumor landed on a market that was already worried about funding.

Memory’s China Problem

Memory stocks fell for most of the week because of one announcement. China’s CXMT said it plans to ship a DDR5 server memory module built on a denser 4F² cell design by the end of the year, Global Sources reported. The current industry standard is 6F². Packing cells into less area lowers cost per bit, which is why the news hit Micron, SanDisk, Seagate and Western Digital on Monday and kept hitting them through Wednesday.

A product plan is a long way from mass production, and none of the memory makers has changed its guidance. The stocks fell anyway. A sector priced for a long upcycle reacted to a prototype announcement as if it were new supply, which says more about positioning than about DRAM. The selling also hit the equipment makers that depend on memory capex: Applied Materials fell −19.3% and Lam Research −17.2%, worse than any memory maker except Seagate.

Lam Research, a major supplier to memory makers, reports on Wednesday after the close, and Western Digital on Thursday.

Inflation Came Back, and the Fed Looked Past It

September’s consumer price index rose 0.6% from August and 3.7% from a year earlier, the fastest annual pace since May, according to the Bureau of Labor Statistics. Energy did most of the work, with gasoline up 9.1% in a month. Core CPI rose 0.3% on the month and 2.6% on the year, with shelter up 0.4% and airline fares up 1.2%.

Producer prices were hotter. September PPI jumped 1.3% and core PPI rose 0.7%, per Zacks. The headline gain was the largest since March 2022. Retail sales rose 0.8% and initial jobless claims stayed below 200,000. On paper, this was a week of strong growth and rising prices.

The rate market moved, then moved back. After the CPI report, the odds of a hike at the October 27-28 meeting jumped from 21% to 54% on CME FedWatch, Benzinga reported. Fed officials then signaled patience, and Chair Kevin Warsh told an IMF event in Bangkok that he doesn’t prejudge meetings. By Friday the October odds were back near 15%. Polymarket still put the chance of a December hike at 79%.

Treasury yields fell through all of it. The 10-year yield ended the week at 5.098%, down 14.6 basis points, its biggest weekly drop since mid-February, before the Iran war began, according to the WSJ. The 30-year fell 16.4 basis points to 5.436%. Yields falling during a week of hot inflation data usually means investors are buying bonds for safety, worried about growth. Here, the growth risk is the AI capex cycle, which has supported much of the economy’s investment spending. If that spending slows because borrowers can’t fund it, the bond market will have been right to look past September’s inflation.

Money Rotated Instead of Leaving

The S&P 500 fell −1.97% for the week, measured by SPY, while the equal-weight version, RSP, rose +0.49%. That 2.46-point gap is the second widest in at least three years. The only wider one came in the week ending June 26 this year, and those are also the only two weeks in that span when the equal-weight index rose while the Nasdaq 100 ETF fell more than 4%.

Breadth told the same story. Consumer staples gained +5.0%, energy +4.9% and health care +3.5%, while technology lost −7.7%. Among companies worth more than $50 billion, the best performers were Accenture (+9.9%), ConocoPhillips (+9.3%), Lockheed Martin (+8.3%) and Northrop Grumman (+8.1%).

Investors sold one theme and bought others without raising cash in a hurry. The June precedent offers a partial guide, with only one case to go on. The Nasdaq 100 ETF bounced 2.3% the following week, but four weeks later it was 3.5% below its June 26 close, while the equal-weight index was 2.3% higher.

Same Drop, Different Fear: Why the VIX Ignored Wednesday and Jumped on Friday

Wednesday and Friday were close to identical days for the S&P 500. It fell −1.11% on Wednesday after the CPI report and the CoreWeave slide, and −1.09% on Friday. At its worst, it was down about 1.45% both days. The VIX reacted very differently. It rose 1.60 points, or +10.5%, to 16.87 on Wednesday. On Friday it rose 3.53 points, or +21.5%, to 19.99, after touching 22.31 during the afternoon.

Over the past three years, the S&P 500 fell between 0.8% and 1.4% on 48 other days, and the median VIX increase on those days was 1.47 points. Wednesday's rise sat almost exactly at that median. Friday's ranked 98th percentile. Only one similar day had a bigger jump: March 6 this year, when the VIX started from almost 24.

So Wednesday was the normal day. Friday was the outlier. Four things set the two days apart.

1. A scheduled shock against an unknown one. CPI came out at 8:30 a.m. on a known date, and traders bought protection beforehand: the one-day VIX, VIX1D, rose 39% on Tuesday. Once the data came out, the uncertainty was gone. Friday's driver was an unconfirmed rumor about a fund in trouble, which came with no size and no end point. The intraday paths show the difference. On Wednesday the VIX peaked at 17.45 at 2:05 p.m., the same five minutes the S&P 500 hit its low. On Friday the VIX peaked at 22.31 at 1:25 p.m., almost two hours before the index bottomed. On Wednesday the stock decline came first and the VIX followed it. On Friday the VIX moved before the stock decline.

2. The usual hedges worked on Wednesday and failed on Friday. On Wednesday, health care rose +2.53%, consumer staples +2.22% and long Treasuries +0.66%. The 10-year yield fell from about 5.30% just after the open to 5.18% by mid-afternoon as stocks slid. A portfolio holding bonds and defensive stocks was partly protected without any options. On Friday, health care fell −0.27%, Treasuries barely moved and the 10-year yield stayed near 5.10% all day. Energy (+2.52%) and utilities (+1.34%) did rise, but energy rose on war risk, which is a reason to buy more protection, not less. With bonds and defensives giving little cover, options were the hedge left.

3. Correlation and the tails. The VIX prices the index, and index volatility depends on how individual stocks move together. Cboe's one-month implied correlation index, COR1M, rose from 7.65 to 10.07 on Friday, a 32% jump, compared with a 19% rise on Wednesday. Over the same days, Cboe's dispersion index, which tracks expected single-stock volatility, barely moved, staying between 36.6 and 37.0 from Tuesday to Friday. Single-stock risk didn't change much, but traders started pricing the chance that stocks would fall together. 24K Research estimated that higher correlation explained about 64% of Friday's VIX move. Demand for crash protection rose faster still. The SKEW index, which prices deep out-of-the-money puts, closed at 168.59, its highest close since February 2025, and VVIX, the volatility of the VIX itself, rose 19% to its highest close since April.

4. Expiration day and the weekend. Friday was October's monthly options expiration. The standard S&P 500 options that many funds use as hedges settled at the open, so anyone who wanted to stay protected had to buy November contracts, the maturities the VIX is built from. SpotGamma's Brent Kochuba told CNBC that traders appeared to be cashing in October VIX positions and rolling into hedges against a larger downturn. One trader bought 9,200 November 20-strike VIX calls for $1.7 million and 12,500 November 23-strike calls for $1.1 million. Add a weekend with Brent near $110 and a war in the Gulf, and Friday afternoon was a bad time to be unhedged. On my own GEX dashboard, dealers are now short gamma. In that position their hedging tends to add to moves instead of absorbing them, so bigger swings in both directions are more likely next week.

Others read the week differently. Saxo's options brief on Wednesday morning argued the index had stayed calm because sector moves were canceling out: "Correlation this low is what has been paying for the calm," it wrote. On Wednesday itself, BigGo Finance suggested the VIX stayed moderate because the S&P 500 was still only about 1.3% below its record high. Both explanations fit Wednesday. They don't explain Friday, when breadth was almost the same: 249 S&P 500 members rose on Friday, against 241 on Wednesday.

Wednesday's VIX looked muted only next to Friday. The useful signals are in Friday's details. The VIX curve flattened: three-month volatility was 18% above the VIX on Tuesday and only 3% above it on Friday, and the nine-day VIX closed above the 30-day VIX for the first time since mid-September. SKEW says traders paid up for tail protection. COR1M, at about 10, is still below its median of the past year. That points to a tail scare inside a market whose stocks still move in many directions. A broad correlation break would push COR1M far higher. October VIX contracts expire on Wednesday morning, and Meta and Tesla report after the close that day.

Banks Opened Earnings Season

The big banks set a mixed tone. JPMorgan earned $5.57 a share, below the $5.92 consensus, as fixed-income trading revenue fell 4% to $5.4 billion, while investment banking fees rose 14%, per Finviz. Bank of America set aside $1.2 billion to cover regulatory fines over failures to detect suspicious activity, and its sales and trading revenue fell 5%, Reuters reported. The financial sector still gained +1.4% for the week, helped by the rotation.

For the S&P 500 as a whole, the season started well. With 9% of companies reporting, 84% beat earnings estimates and blended third-quarter earnings growth stood at 29.0%, according to FactSet. The forward price-to-earnings ratio is 19.0, below its five-year average of 19.8.

Oil Kept Its War Premium

The November WTI contract settled at $101.41 on Friday, up +10.41% for the week, and December Brent at $109.86, up +4.91%. Attacks on shipping around the Strait of Hormuz and on Saudi energy sites kept supply risk in the price, as Friday's recap detailed. After Friday's close, President Trump waived most of a new Russia sanctions law, part of a push to keep energy prices down before the midterms, the WSJ reported. Oil traders will get their first chance to price that on Sunday evening. Energy stocks were the week's second-best sector. Tudor Jones, in the same interview, said he is long oil and expects another $20 to $25 a barrel after the midterms. Oil at these levels feeds straight into the inflation numbers above, which is why the Fed can't fully look past them for long.

My Take: A Bounce, Then Another Leg Down

I don't think this ends as a clean, healthy correction. My base case is more downside, though not in a straight line: a short-term bounce, another leg lower, and only after that the real rally.

Chip stocks' own history is part of why. Since the iShares Semiconductor ETF (SOXX) began trading in 2001, it had fallen for seven straight sessions and then opened more than 1% lower only twice before this week: in September 2002 and October 2008. Both times the ETF bounced within four sessions, by about 6% in 2002 and 19% in 2008. Both times it then fell to a lower low within weeks, 20% below the bounce high in 2002 and 37% below it in 2008, and those lows turned out to be the bottoms of each chip bear market. The same signal fired again on Thursday. Two cases is a thin sample, and both came in much deeper bear markets than this one. I still think the shape fits: a sharp rebound that doesn't hold, then the real low.

That leaves me with two scenarios. In the first, a strong rally comes within a week or two and then gives way to another decline. In the second, the market keeps sliding into the midterms, rallies hard from early-to-mid November, and then corrects again. Paul Tudor Jones's call fits the second one better: weakness now, a bottom in late October or early November, a rally into year-end, and the real trouble in 2027. The calendar argues for a bounce somewhere in this window. In the 19 midterm years since 1950, the S&P 500 rose from Oct. 16 to the end of the year 15 times, by 4.9% on average, and it was higher 12 months later every time. If the bounce comes late, I'd look for it right after the Nov. 3 vote.

This selloff also looks different from July's. Then, AI hardware collapsed while software rallied. From the SOX's June 22 record to its July 29 low, the index fell −28.6% while Salesforce and ServiceNow each gained about +25%, as funds that had bet against software were squeezed. This time software is no shelter. The iShares software ETF fell −7.4% this week, dragged down by Oracle (−9.3%) and Palantir (−11.0%), even though Salesforce rose. Oracle never really recovered from July: it lost −32.7% into the July low and won back only +20.1% in the rebound.

What I see underneath is a split by balance sheet. Companies that need outside money and don't generate much cash are being sold as long-term yields rise and credit gets more expensive. Microsoft rallied +37.0% from the July low to Oct. 9 and lost only −1.1% this week, and Meta looks similar. The hyperscalers need enormous amounts of capital too, but they also produce enormous cash, so money is crowding into them. The Wall Street Journal made the same point this week, quoting AllianceBernstein's Ali Bendarkawi: "the marginal borrower in the AI trade is anyone who's not a high-quality hyperscaler." The catch is that the hyperscalers are tied to everyone else. If data centers, memory and power keep falling, confidence in the big names will crack as well, and I expect them to correct too.

In the end it comes back to inflation, which means oil. The Iran war shows no sign of ending, and CNN reported that the Pentagon has drawn up plans for new strikes on Iranian energy and nuclear sites, possibly including the largest conventional bomb in the U.S. arsenal, per Ahram Online. Washington is now easing pressure on Russia to bring diesel prices down, and even Republicans such as Ted Cruz have objected. It may be framed as the national interest, but a fight like that right before an election can leave investors more confused. France adds to the pressure on long-term yields: its 10-year yield touched its highest level since 2002 this month, and cost-of-living protests turned violent across French cities on Saturday, the WSJ reported. Some investors see Treasury yields at these levels as a buying opportunity. Then oil, the war or another overextended borrower changes the picture, and demand for safety shows up one day and disappears the next. It's confusing.

So I'm watching two things: long-term Treasury yields and the hyperscalers' stock prices. If the Magnificent Seven hold up and the 10-year and 30-year stop climbing, I think this ends as a shallow decline followed by a short rally. The 10-year has already come down from 5.31% on Oct. 5 to 5.10%. The angle of the drop is unusual, though. The Nasdaq 100 lost −6.5% in eight sessions from its Oct. 6 record, nearly two-thirds of what it lost over about five weeks in July, and that speed is one reason I expect more downside. The S&P 500 closed Friday at 7,659, below its 50-day moving average near 7,706 for the first time since Sept. 16. December E-mini futures dipped to 7,676 and finished the day near 7,704, so 7,700 held in futures, barely. If this zone holds, dip buyers could show up again the way they have all year. Where it goes next week is the open question.

The Week Ahead

The Fed is in its quiet period before the October 27-28 meeting. Earnings are the main event, with Meta and Tesla reporting on the same evening as the VIX expiration. All times are Eastern, and company schedules are from StockTitan's earnings calendar.

Monday, Oct 19

• 10:00 a.m.: NAHB homebuilder sentiment
• Earnings: Fifth Third (before the open), Steel Dynamics, W.R. Berkley (after the close)

Tuesday, Oct 20

• 8:30 a.m.: Housing starts and building permits
• Before the open: GE Aerospace, GM, RTX, 3M, Northrop Grumman
• After the close: Netflix, Capital One, Intuitive Surgical, and SpaceX's first report as a public company, with a webcast at 4:30 p.m.

Wednesday, Oct 21

• Morning: October VIX futures and options expire
• 1:00 p.m.: 20-year Treasury bond auction
• Before the open: Thermo Fisher, AT&T, Abbott, NextEra Energy
• After the close: Meta, Tesla, Texas Instruments, Lam Research, IBM, SAP

Thursday, Oct 22

• 8:30 a.m.: Initial jobless claims
• Before the open: Procter & Gamble, Lockheed Martin, Union Pacific, Honeywell, Blackstone, Comcast
• After the close: Western Digital, Bloom Energy, Newmont

Friday, Oct 23

• 9:45 a.m.: S&P Global flash PMIs for October
• 10:00 a.m.: University of Michigan consumer sentiment (final)
• Before the open: American Express, SLB

Lam Research and Western Digital will show whether memory customers are cutting orders after the CXMT news. Bloom Energy reports in the middle of the debate over who funds AI power. Meta's capex guidance matters more than its earnings this quarter.

Bottom Line

The AI selloff this week was about financing. Chip demand, foundry orders and equipment backlogs all came in strong, while the companies that borrow to build AI capacity saw their bonds trade like distressed debt. Stocks followed the credit market.

The rest of the market held up. Equal-weight stocks rose, defensives and energy rallied, and the S&P 500's earnings are still growing close to 30%. The VIX puzzle shows where the fear sits. A hot inflation report was a known risk that bonds and defensives could absorb. A possible forced seller was not, and traders paid for protection on Friday afternoon.