The $3 trillion ThursdayThursday's tech sell-off, and what it means if you build AI products

Every one of the S&P 500’s eleven sectors closed lower, and by mid-afternoon 419 stocks in the index were down against 84 up. The Dow lost more than 1,800 points, its worst second session after a midterm election in records going back to 1896, and the Russell 2000 small-cap index joined the Nasdaq in correction territory. The VIX, the market’s fear gauge, jumped more than 60 percent to around 27.

The damage was lopsided in a way that matters. Nvidia fell about 9 percent and is back below where it ended last year. AMD and Micron lost about 15 percent each, Broadcom about 10 percent and Arm about 12 percent. CoreWeave, which borrows to build GPU capacity and rents it out, fell about 9 percent on top of Wednesday’s drop, roughly 18 percent in two days. Meanwhile Alphabet slipped 1.6 percent, Microsoft 2.8 percent and Meta 3.4 percent.

The market sold the companies that build and finance AI capacity much harder than the companies that own it. One way to read that, which several strategists made on Thursday, is that if new capacity gets harder to build, the capacity that already exists becomes more valuable.

On the $3 trillion figure doing the rounds tonight: it is for one session, and the count depends on which index you use. Fortune put the Nasdaq’s loss at more than $3 trillion at midday, when the index was down 5.2 percent, and the close was worse. Russia’s economy produces about $2.9 trillion a year on the IMF’s latest estimate, so the comparison people are making on X tonight holds, narrowly. I will update this paragraph when the closing tally is published.

Retail investors, who bought every dip for two years, sold. Vanda Research counted $358 million of net selling in individual stocks by 3 p.m., on track for the biggest retail selling day since 17 March 2020, and called it “a total capitulation by retail investors”.

Tuesday’s midterms gave Democrats both chambers of Congress. On its own that did not frighten anyone; the Dow rose 1,569 points on Wednesday. What changed the maths for AI was what happened lower down the ballot.

Voters in ten states faced close to 40 local measures on data centres, and by Capital B’s count 77 percent of them voted against development. Forbes’s list runs to 41 measures, with Ohio alone passing 13 outright bans, and depending on how you count, between one and four local votes went the industry’s way. Ohio’s governor-elect, Amy Acton, campaigned on a conditional moratorium on new data centres unless they cover their full utility costs, and Georgia’s Keisha Lance Bottoms backed a pause on new construction. Texas, where a large share of the planned capacity was meant to go, had already paused permitting in September.

The industry heard it. Meta’s president, Dina Powell McCormick, said on Thursday that building data centres in the US is “going to be harder” and that hyperscalers are weighing whether to build outside the United States. Dan Ives of Yorkville, one of the loudest AI bulls on the street, now expects capital spending growth to peak next year instead of 2028 or 2029, and estimates that “on the low end 10% of the planned 800-1,000 AI focused data centers in the US will now be wiped off the roadmap”.

If your 2027 revenue depends on sites that may never be permitted, you got repriced today. That describes memory makers, power-equipment makers and the GPU clouds that borrowed against future capacity, and that is exactly the list of stocks that fell hardest.

A political headline cannot take 5 percent off the Nasdaq by itself. It needs a market that was already looking for a reason to sell, and October supplied three.

1. Mon 6 OctNasdaq Composite closes at a record 27,599.89, the peak this correction is measured from.
2. Wed 7 OctThe 10-year Treasury yield trades above 5.35 percent, its highest since 2002.
3. Thu 8 OctReports put OpenAI’s annualised revenue around $50 billion, well below earlier figures. Nvidia, Oracle and CoreWeave fall.
4. Thu 15 OctCoreWeave’s default insurance trades above 1,100 basis points, a distressed level. Oracle’s hits a record.
5. Wed 28 OctThe Fed holds at 3.75 to 4.00 percent, with two members voting for another increase.
6. Thu 29 OctAlphabet falls about 13 percent, its worst day on record, after reporting negative free cash flow and keeping 2026 spending at $195 to $205 billion.
7. Fri 30 OctAmazon reports negative free cash flow on about $220 billion of spending. The Nasdaq ends its worst month since March 2025.
8. Tue 3 NovElection day. After the close, AMD guides above expectations and says demand still exceeds supply.
9. Wed 4 NovDemocrats win both chambers. The Dow rises 1,569 points. Underneath, Nvidia and CoreWeave already fall.
10. Thu 5 NovThe Nasdaq erases its post-election gain in six minutes and closes down 5.5 percent.

The first reason is cash flow. In the earnings season that ended last week, Amazon, Alphabet and Meta all reported negative free cash flow, which is the cash a business has left after paying for its operations and its building programme. Amazon’s figure was minus $42 billion, Alphabet’s minus $9 billion and Meta’s minus $2 billion, and Microsoft’s free cash flow fell 27 percent to $19 billion, with a plan to go to the bond market for more. Alphabet lost about $500 billion of value in a single day on 29 October for keeping its spending guidance at $195 to $205 billion. BlackRock’s numbers show why investors flinched. Hyperscaler capital spending has risen to about $650 billion over the past year from about $360 billion, while the margin earned on each extra dollar of revenue has fallen to about 28 percent from 40 percent.

The second is credit. CoreWeave’s credit default swaps, which are the price of insuring its debt against default, traded above 1,100 basis points in mid-October, a level the market reserves for companies in distress, and its 2031 bond yielded almost 16 percent. Oracle’s swaps hit a record 288 basis points. The AI build-out is increasingly financed with borrowed money, and the lenders are now charging for the risk. Oracle eased some of that pressure over the weekend by capping its new borrowing at $20 billion through 2030.

The third is the price of money itself. The 10-year Treasury yield traded above 5.3 percent in early October, its highest since 2002, after the Fed raised rates in September. The Fed held in October with two members voting for another increase, and prediction markets put the odds of a December hike near two in three. AI infrastructure is a long-duration bet, and long-duration bets are the first thing investors sell when rates rise.

Add positioning. Nomura’s Charlie McElligott summed up Thursday as “they’re shooting the generals”, meaning hedge funds locking in a good year by selling the stocks that made it, with momentum funds selling on top of them.

Expensive money, negative cash flow and a political cap on supply is a bad combination for any business whose plan is to borrow, build and then find the tenant. It is a much smaller problem for a business that funds its build-out from a profitable core and can slow down when it wants to.

This is the part of Thursday I keep coming back to. Nothing in the demand for AI changed between Wednesday and Thursday.

AMD reported after Tuesday’s close, guided fourth-quarter revenue above what Wall Street expected and said demand still exceeds its supply. Lisa Su, its chief executive, said on Thursday that the market can be “a little bit short-sighted”. Microsoft trimmed its capital spending guidance last week because it cannot get GPU capacity from third parties fast enough, which is a shortage. Anthropic’s updated listing filing shows revenue of $17 billion in the third quarter, up 47 percent on the quarter before. And the wider economy is fine. Announced layoffs in October were the lowest since 2023, and S&P 500 earnings are growing at close to 40 percent year on year by Apollo’s count.

In my own work nothing moved either. The enterprise buyers who asked us for payback numbers last week will ask harder next week, and that, rather than any stock price, is how Wall Street reaches a company like mine.

I find it more useful to put numbers on scenarios than to argue about whether this is a bubble. These odds are my judgement, not market prices, and I will revise them in public as the data comes in.

What separates the middle case from the worst is credit, so that is where I am looking.

Why I do not put the dot-com case higher. In 2000 the index was barely earning and the whole build-out ran on new equity. Today S&P 500 earnings are growing fast, and the hyperscalers fund most of their spending from profitable core businesses. The 2000 comparison holds for the hardware and GPU-cloud layer, where spending runs ahead of revenue and is borrowed at the edges. For the index as a whole, the closer analogues are late 2018 and 2022, a rates-plus-positioning correction in an expensive, concentrated market.

And it is expensive. The Shiller CAPE, which compares prices with ten years of inflation-adjusted earnings, is near 41, higher than at any time outside 1999 to 2000 and this year, and the ten largest stocks make up close to 39 percent of the S&P 500, with Nvidia alone around 7 percent. When one stock is 7 percent of the index, a bad day for that stock is a bad day for everyone’s pension.

Thursday’s drop is the kind of day that starts a lot of charts, so it helps to see where a 12 percent Nasdaq correction sits next to the real drawdowns of the past 26 years.

Two things stand out. Days of 4 percent or worse tend to cluster in the middle or at the end of a drawdown rather than at the start, and a VIX of 27 is nowhere near the 65 it touched in August 2024 or the 80-plus of March 2020. Thursday’s retail selling was the biggest since March 2020, which is the sort of thing you see near bottoms, but one day of it is not a bottom. My working assumption is that the washout is still ahead of us, and that it is a chance to buy time and capacity rather than a reason to panic.

These are the four decisions I am taking into Monday.

1. Treat compute as a buyer’s market that has not arrived yet. Capacity is short today. Microsoft’s GPU delays run into the first half of 2027 and AMD says demand exceeds supply. If permits slow new US sites, existing capacity gets pricing power for a while, and if GPU clouds fail to refinance, their capacity gets dumped later. So I am keeping GPU contracts short and flexible, spreading them across more than one provider, and reading each provider’s balance sheet. A provider whose default insurance trades above 1,000 basis points is a counterparty risk as well as a vendor.
2. Raise earlier and expect a markdown. Private AI valuations follow the public ones with a lag of a quarter or two. Anthropic’s expected IPO valuation has already been cut to about $1.5 trillion from about $2 trillion, and the first private casualties will be in infrastructure. Application companies with revenue will get funded, at a lower price. I would plan for six to twelve more months of runway than you thought you needed.
3. Sell payback. The market punished AI spending that cannot show a return, and enterprise buyers will copy that within a quarter. Lead every conversation with the measured payback, and price on usage so the buyer’s risk falls with yours.
4. Hire into the dip. The labour market is in a low-hire, low-fire pattern. If public AI-infrastructure companies tighten budgets, senior people who were unreachable six months ago become reachable. I wrote about what AI is doing to tech jobs; this is the other side of that story.

• Fri 8:30 a.m. ETThe October jobs report. Consensus is 59,000 jobs and 4.2 percent unemployment. A big miss turns a sector story into a recession story.
• All weekWhether CoreWeave’s and Oracle’s default insurance stabilises. That is the line between my middle case and my worst case.
• Later in NovAnthropic’s IPO pricing. A strong debut confirms private AI valuations; a weak one chills late-stage rounds into 2027.