For two years, the AI trade had one rule: build more, and the market rewards you. On Thursday, November 5, that rule got its first real stress test.
The Nasdaq Composite fell 5.52%, its worst day in 19 months. It closed 11.9% below its October 6 record, which puts it in correction territory. The S&P 500 fell 4.36% and the Dow 3.44%. Roughly $3 trillion in market value was wiped off S&P 500 companies in one session.
The falls were steepest exactly where the AI build-out money flows: chips, memory, and the data centre operators.
Two patterns stand out. First, the closer a company sits to physical data centre construction, the harder it fell. Memory makers and neoclouds lost 10–22%; Amazon and Alphabet lost about 2%. Second, it went global: Japan’s Nikkei 225 and Korea’s Kospi, home to SK Hynix and Samsung, both closed about 6% lower overnight.
A few more numbers show how unusual the day was:
- The Nasdaq 100 fell 6.36%, its worst day since March 2020. It’s now 14% below its high and broke below its 200-day moving average, a widely watched support level.
- The iShares Semiconductor ETF fell 11%, its worst day in six years.
- Retail investors sold more than $330 million of single stocks, their biggest day of selling since the pandemic crash, per Vanda Analytics.
- The day before, Wednesday, the Dow had its best day since April 2025. Thursday erased all of it.
The trigger: data centres were on the ballot, and they lost
Here’s the odd part. Markets rallied the day after the November 3 midterms. The result was clear and uncontested, and oil fell on hopes the Iran war would wind down.
The rally lasted one session. Then investors read the fine print: in many states, data centres themselves were on the ballot.
- 39 of 40 data centre ballot measures with complete counts went the restrictive way, per Redbud Consulting’s tally.
- At least 38 jurisdictions approved bans or restrictions on data centres.
- Democrats won the governorships of Texas and Ohio, two of the busiest data centre markets.
Texas is the big one. It holds roughly a quarter of all planned US data centre capacity. The governor-elect has pledged to pause new data centre approvals until there’s legislation protecting residents and electricity ratepayers.
Why do voters care? The objections are concrete: higher power bills, heavy water use for cooling, land use, and very few permanent jobs once construction ends.
Why do markets care? Because a data centre needs four things: money, chips, land and a grid connection. For two years, investors only priced the first two. A permit pause in a state with a quarter of planned capacity hits the other two directly. Projects can move to friendlier states, but new sites, power and approvals take time, and for a debt-funded project, every month of delay is interest paid with no revenue coming in.
The selling picked up through Thursday afternoon after senior lawmakers publicly called for an immediate moratorium on data centres. Jefferies also cut its forecast for US data centre capacity through 2030.
The projects at stake are not small. Meta alone has a $50 billion data centre (Hyperion) under construction in Louisiana and a $14 billion one in Texas.
How much money is riding on the build-out
To see why a permit fight can move trillions in market value, look at what the four biggest spenders plan to pour into infrastructure this year.
Amazon, Alphabet, Microsoft and Meta guided to roughly $735 billion of capital spending in 2026, up about 79% from around $410 billion in 2025. Most of it goes to AI data centres, GPUs, custom chips and power.
Every one of those dollars is someone else’s revenue: chipmakers, memory makers, server and networking firms, power equipment suppliers, and the neoclouds that rent out compute. That’s why a threat to the pace of construction hits the whole supply chain at once, not just Big Tech.
It also matters for the wider US economy. In September, data centre construction spending jumped 80%, while spending in most other construction categories fell. When one category is carrying that much of the growth, a slowdown there shows up in GDP, not just in tech stocks.
The other side of the ledger is the revenue needed to justify it. Estimates put the annual AI revenue required to pay off this infrastructure at more than $2.5 trillion by 2030. That has to come from customers using AI to raise their own profits, mostly by cutting costs.
Where the stress actually sits
Not every AI stock carries the same risk. Three numbers explain where the fragility is.
1. The market has never been this concentrated in one theme
In 2008, tech, media and telecom made up about one-fifth of the US large-cap equity market. Today it’s nearly half. When half the market rides on one story, a wobble in that story moves everything, including your index fund.
2. A lot of AI money is AI money going round in circles
This is circular financing: Company A invests in Company B, Company B spends that money on Company A’s products, and Company A books it as revenue.
- Nearly 29% of AI firms’ investment deals are in other AI firms.
- More than half of the investment coming into AI firms comes from other AI firms.
- Microsoft has invested more than $13 billion in OpenAI, which buys cloud capacity from Microsoft’s Azure.
- Meta has signed about $41 billion of compute contracts with Nebius and CoreWeave.
None of this is illegal. But it means some of the demand propping up the sector comes from inside the sector. The same loop ran in the late 1990s as “vendor financing”, when telecom equipment makers lent customers the money to buy their own gear. It worked until the boom stopped.
3. The debt is concentrated in the neoclouds
The hyperscalers fund most of their spending from cash flow. The neoclouds, firms like CoreWeave, IREN, Nebius and Applied Digital that build data centres to rent out compute, lean much more on borrowing. A permit delay hurts them most, because their interest bill doesn’t pause.
The market has already noticed:
- CoreWeave is down about 70% from its May peak.
- Intel fell nearly 30% in one day last week and another 10% on Thursday, more than halving in a month.
- The Nasdaq had already fallen about 5.3% in October, before the midterms.
Is this 2000 again? The data cuts both ways
The internet was real, and it still crashed markets 25 years ago. A technology can change the world and still be a bad investment at the wrong price. So the useful question isn’t “is AI real?” but “how does the money look compared with then?”
| Dot-com bust (2000–02) | AI boom (2026) | |
| Who’s doing the spending | Many young firms with little or no revenue | Mostly the four largest, profitable tech companies |
| How it’s funded | Heavy debt and vendor financing | Mostly cash flow; heavier debt at neoclouds |
| Circular deals | Vendor financing common | ~29% of AI firms’ deals are in other AI firms |
| Tech’s weight in the market | Large | Nearly half of US large caps (vs one-fifth in 2008) |
| Damage to infrastructure names | Nortel and Lucent lost ~99%; Global Crossing and WorldCom went bankrupt | CoreWeave down ~70% from peak; no major defaults yet |
| Where the index stands | Nasdaq fell ~78% peak to trough | Nasdaq down ~12% from peak, still up ~4.6% in 2026 |
The left column is why people are nervous. The right column is why this isn’t 2000 yet: the biggest spenders are profitable, and the damage so far is concentrated in the most leveraged corners.
The warning sign from the dot-com era worth tracking: booms end when lenders stop lending. Defaults and pulled credit lines among debt-funded builders would mark the shift from a repricing to something worse.
The offsets
Not every number this week was bad.
- Oil is falling. Crude dropped below $100 a barrel, from $118 at the end of October. Brent fell as much as 20% this week. Cheaper energy eases inflation, which helps rate-sensitive sectors like housing, REITs, utilities, airlines and consumer businesses.
- The year is still positive. Even after Thursday, the Nasdaq is up about 4.6% in 2026.
- Buyers showed up fast. By Friday’s pre-market, some of the hardest-hit US stocks were up around 4%.
- Demand hasn’t vanished. Intel reported a 78% jump in data centre revenue last quarter, and AMD said its data centre sales more than doubled.
One offset works the other way: Federal Reserve officials signalled on Thursday that further rate hikes are possible. Higher borrowing costs bite hardest on exactly the debt-funded projects at the centre of this sell-off.
What to track next
Whether this becomes a deeper correction or a sharp repricing will show up in a handful of numbers:
- Permits. How many data centre projects get paused, and how much of Texas’s planned capacity is affected.
- Capex guidance. Whether the big four hold, raise or trim their roughly $735 billion 2026 plans, and what they guide for 2027.
- Chip orders. Cloud providers’ orders for GPUs and memory. A slowdown, or signs of spare capacity, would confirm overbuilding fears.
- Neocloud credit. Bond spreads, refinancing and any pulled credit lines at the debt-funded builders.
- Rates and oil. A sustained fall in oil and inflation supports stocks; further Fed hikes work against them. Friday’s October jobs report is the next data point on the Fed’s path into 2027.
The bottom line
So, are we in an AI crash? By the numbers, not yet. A 12% drop from the peak is a correction, the index is still up for the year, and no major company has defaulted.
But the question the market is pricing has changed. For two years it was how fast can this be built? Now it’s where can it be built, how long will approvals take, and will $735 billion a year ever earn its keep?
The data points one way: the profitable giants with capacity already in place can absorb a slower build-out. The leveraged builders can’t, and that’s exactly where the losses have piled up.
This article is for educational and informational purposes only. It is not investment advice or a recommendation to buy or sell any security. Investing in global markets involves risks, including currency risk and loss of capital. Please do your own research or consult a qualified advisor before investing.