AI's borrowed boom: why 66 AI build-out stocks have lost about $2 trillion

The short version

• In our screen of US-traded stocks that fell 40% or more from their highs of the past six months, 66 are AI build-out landlords and suppliers (we leave out most chipmakers, which mostly fell with their sector). Together they have lost about $2 trillion (our math) as the borrowed money behind the build-out got costlier and harder to get, and doubts about returns grew.
• Until 2025, the tech giants paid for AI mostly from their own cash. In 2026 spending outgrew the cash some of them generate: across the five biggest spenders, new debt covered 32% of capex, up from 9% in fiscal 2024 (FactSet).
• Oracle, the “neoclouds” CoreWeave and Nebius, and former bitcoin miners such as IREN borrow far more, against contracts with a handful of AI customers.
• Lenders grew warier and doubts about returns spread; suppliers fell too, even with rising orders.

Why it matters now

Over the past 18 months the AI build-out came to depend on outside money, and in 2026 that money got dearer and harder to get. The first blow came in early June, when a record share sale and new borrowing plans showed how much funding the build-out needs; the second, from mid-September, came as credit markets turned on the most indebted builders. The key moments:

• Jun 1–10: Alphabet announced an $80 billion equity raise (upsized to $84.75 billion), the largest US follow-on offering on record according to LSEG, and Oracle said it would raise about $40 billion in fiscal 2027. Half of the stocks had peaked by Jun 2, and most (41 of 66) peaked between late May and the end of June.
• Sep 15–16: the 10-year yield closed at 5% for the first time since 2007, and the Fed raised rates to 3.75–4.00%, its first hike since 2023, as oil prices from the Iran war kept inflation elevated.
• Oct 12: Broadcom CEO Hock Tan said capital “might be the biggest constraint” on AI growth, ahead of chips or power.
• Oct 13–16: CoreWeave’s bonds fell to about 77–80 cents on the dollar and its credit default swaps topped 1,000 basis points. Gartner published its impairment warning on Oct 14.
• Oct 21–22: Meta’s AI spending pushed its free cash flow below zero (our Meta article), and AI power stocks fell with it (GE Vernova).

Rates aren’t the whole story: the Nasdaq hit a record on Oct 6 with the 10-year above 5.2%, and the sharpest losses came after yields started falling. Some stocks have their own problems, such as new memory supply and MasTec’s weak cash flow (−19% in a day).

How borrowed money funds the AI build-out

The chain works only while each link can keep raising money. AI labs such as OpenAI rent computing power from landlords (the hyperscalers, Oracle, neoclouds and miners-turned-hosts), which borrow to buy chips and build data centers; suppliers depend on those orders. Every link is valued on profits expected in 2027–2028, so the cost and availability of credit ripple down the chain.

The landlords’ squeeze

The landlords’ problem is timing: interest comes due now, while the revenue meant to repay the debt arrives over years. CoreWeave is the clearest case:

• At the end of June it had about $35 billion of debt and $5.5 billion of cash.
• Second-quarter interest expense was $640 million, against $128 million of adjusted operating income. It guided third-quarter interest to $860–940 million, several times its operating profit.
• D.A. Davidson’s Gil Luria estimates that CoreWeave borrows at about 9% while earning returns near 1%.
• Bonds it sold at 9.625% now yield about 15%.

Nebius had about $8.5 billion of convertible debt and $8 billion of cash at the end of June, plans $20–25 billion of capex this year, and gets 59% of its revenue from three customers. On Oct 22 it signed a $3.5 billion, one-year secured loan from Citi at 2.75 points over short-term rates. IREN plans $25–30 billion of capex in fiscal 2027, part of it at a 9% fixed rate.

Oracle: a strong business, stretched by the build-out

Oracle has one of the biggest AI contract backlogs, but its spending now far exceeds the cash its business generates, so it leans on debt and stock markets. Its contracted backlog reached $664 billion at the end of August. In July, S&P estimated that OpenAI accounted for roughly half of the $638 billion backlog Oracle had then. The strain shows in its finances:

• Free cash flow was −$23.7 billion in fiscal 2026 and −$5.4 billion in the quarter to August.
• It raised $43 billion of debt and $5 billion of equity in fiscal 2026, then sold $20 billion more of stock.
• S&P cut its rating to BBB−, one notch above junk, on July 9.
• Its five-year credit default swaps reached 315 basis points on Oct 19, about four times its level a year earlier and far above typical investment-grade levels (CNBC).

How long a GPU lasts

How fast the chips wear out decides whether today’s profits are real. The big clouds and CoreWeave depreciate servers over five to six years, Nebius over five; investor Michael Burry argues the real life is two to three. FactSet notes that prices for older Nvidia chips are holding up, which so far supports the longer schedules.

Strong demand, uncertain returns

Demand is not the problem; the gap between spending and revenue is. Price trackers put rentals of Nvidia’s newest B200 chips near $8 an hour in late September, up from about $5–6 over the summer (Ornn index). But Gartner expects nearly $1 trillion of AI infrastructure spending in 2026 and $6 trillion more through 2030. Covering that, plus financing and a 12% return, would take about $8.5 trillion of revenue, while Gartner forecasts at most $3.6 trillion of AI revenue.

Who’s affected

The losses run along the whole chain, in four groups (our math): the borrowing landlords fell a median 55% (about $706 billion lost), optics and components makers about 50% ($629 billion), memory and storage makers 46% ($404 billion), and builders and power-equipment makers 57% ($302 billion). The common thread: all depend on the build-out continuing at full speed.

Moves are from each stock’s highest close since April 23 to the Oct 22 close; Fujikura and Furukawa use their Tokyo listings.

Within the landlords, the miners-turned-hosts, the smallest, fell hardest. Memory makers also face their own price cycle, as Samsung, SK Hynix, China’s CXMT and Toshiba add capacity. Optics makers mostly fell without bad news of their own; Nokia’s AI and cloud orders rose 118% in the third quarter.

The sell-off is global. On Oct 15, the Tokyo session after CoreWeave’s bond slump, fiber makers Fujikura and Furukawa Electric fell about 13% and Sumitomo Electric 9%, with no news of their own; Kabutan’s market report blamed CoreWeave’s bonds and higher AI funding costs. That came even though TrendForce expects one widely used fiber type to cost nearly five times last year’s price this year. Spain’s ACS, whose US unit Turner builds data centers, is down about 38% in Madrid.

What’s rising instead: cash-generating software that uses AI (CrowdStrike, Snowflake and Datadog are each up more than 100% from April lows), oil refiners and Treasuries.

What the market and analysts think

Stock analysts are split, and credit markets are more worried than most of them.

The bear case

• HSBC rates CoreWeave Reduce ($39 target), estimating it will earn about $12 million per megawatt in 2027, against $17–18 million for the big clouds.
• S&P cites Oracle’s dependence on OpenAI.
• At their mid-October peak, CoreWeave’s credit default swaps implied more than a 60% chance of default within five years, Bloomberg reported, using ICE data.

The bull case

• Goldman Sachs upgraded CoreWeave to Buy on Oct 19, while cutting its target to $106 from $139. It argues that CoreWeave’s borrowing costs should rise only 1–2 percentage points.
• Hock Tan, despite his capital warning, says demand will exceed supply through 2029 and possibly 2030.
• Mizuho expects NAND flash prices to stay elevated through 2027.

A stress test

Dario Perkins of TS Lombard modeled a hypothetical AI bust in a scenario analysis on Oct 22, as reported by Seeking Alpha. If AI capital spending collapsed, he estimated, US stocks could lose about $20 trillion of market value, with the S&P 500 down 30% and tech stocks 40%, and credit losses could reach $125–190 billion. He also wrote that “as far as recessions go, we don’t think this one would be particularly severe,” and that “we don’t know exactly when this AI capex splurge will end.”

Valuation

Oracle trades at about 15.5 times its guided fiscal 2027 adjusted earnings (our math), below the S&P 500’s 19 times (FactSet, Oct 16); the neoclouds are valued more by credit markets.

Have we seen this before?

Yes. Sell-offs concentrated in one debt-funded or capital-hungry corner of the market have happened several times, and the outcome depended on whether demand kept growing and whether the borrowers could keep raising money.

Recent cases first

• In 2022, as bitcoin fell 77% (November 2021 to November 2022), the miners Marathon and Riot lost about 96% from their 2021 peaks. The lender Celsius went bankrupt in July 2022, and Core Scientific, which hosted Celsius’s mining rigs, followed on December 21 with about $4 million of liquidity. Bitcoin regained its high in March 2024; both miners still trade more than 80% below their peaks. Core Scientific relisted in January 2024 and, now an AI host, is one of our 66.
• In 2021–22, profitless growth stocks collapsed: the ARK Innovation fund, a basket of cash-burning tech companies, fell 81% from February 2021 to December 2022 as interest rates rose, while the S&P 500 lost 4% over the same span. It is still about 48% below its peak.
• In 2018, memory chips slumped: Micron fell 54% from May to December as DRAM prices slid; it later cut output. The S&P 500 fell 13%; Micron regained its high in November 2020.
• In 2014–16, debt-funded US shale drillers met an oil-price collapse: the SPDR oil and gas exploration fund fell 72% from June 2014 to February 2016, while the S&P 500 slipped 2%. Junk-bond spreads for oil and gas companies peaked near 1,700 basis points (Markit), and 114 North American producers went bankrupt in 2015–16 with about $74 billion of debt (Haynes and Boone).
• In 1997, the Asian crisis knocked the Nasdaq down 12% from its October 9 peak to its October 27 close, the day the Dow dropped 7%, and 14% by Christmas Eve. It regained its high in February 1998 and then rose another 189% to its March 2000 peak. Dan Ives of Yorkville Ives (formerly of Wedbush) told Yahoo Finance on Sep 29 that “we’re only 15% through with the spending of the $4 trillion to $5 trillion that’s going to happen. I still believe this is a 1997 moment, not a 2000 bubble moment.” (In 1997, the dot-com boom still had more than two years to run.)
• In the 2000–02 dot-com and telecom bust, the Nasdaq fell 78% over 31 months, in stages: down 37% by May 2000, up 35% by July, then far lower. Corning, JDS Uniphase and Ciena lost 98–99%, and the debt-funded carriers Global Crossing and WorldCom went bankrupt in 2002, when only 2–5% of US fiber was lit (FDIC). Even Cisco fell 89%. The Nasdaq regained its peak only in April 2015.

Episode	What sold off	How far	Broad market	What happened next	Back to the peak
2022	Bitcoin miners	−96% (bitcoin −77%)	S&P 500 −20%	Celsius, Core Scientific bankrupt	Bitcoin Mar 2024; miners never
2021–22	Profitless growth (ARKK)	−81%	S&P 500 −4%	Rate hikes cut off cheap funding	Never
2018	Memory (Micron)	−54%	S&P 500 −13%	Prices fell, output cut	Nov 2020
2014–16	US shale drillers	−72%	S&P 500 −2%	114 producer bankruptcies	Never
1997	Nasdaq (Asian crisis)	−14%	Dow −7% in a day	Boom ran 2+ more years	Feb 1998
2000–02	Nasdaq; optical gear	−78%; −98% to −99%	S&P 500 −49%	Carriers bankrupt	Nasdaq Apr 2015; Ciena never
2026 so far	Our 66 build-out stocks	−53% (median)	Nasdaq −4%, S&P 500 +1%	?	—

What separated the good cases from the bad

In the cases that recovered (1997, memory in 2018, bitcoin itself), demand kept growing and the main buyers weren’t forced sellers or bankrupt borrowers; in memory, output cuts also cleared the glut. The bad cases had the opposite: borrowers that couldn’t refinance (shale, crypto miners), cash-burning companies cut off by higher rates (2021–22), and buyers that went bust themselves (2000–02).

Today has features of both. Demand and the biggest buyers look like the good cases: rental prices for the newest GPUs rose through September, and four of the five biggest spenders carry debt of about one times EBITDA or less (FactSet). The borrowed layer looks like the bad ones: CoreWeave’s bonds yield about 15%, and Jason Teh of Clime Investment Management notes that Nvidia’s guarantees and capacity commitments equal about 22% of its forecast revenue, against 13% of Cisco’s and 21% of Lucent’s actual revenue in 2000. So far the stress is concentrated in that layer: from June 2 to October 22 the Nasdaq slipped 4% and the S&P 500 rose 1%, and both set records on October 6.

Views differ. Citadel’s Ken Griffin said on October 20 that there are “almost certainly bubbles within this moment in time.” Teh says investors are “paying for earnings, not for hope.” Gartner expects more than 80% of neoclouds to lose their independence by 2030. Which path this one takes should show first in credit, in GPU rental prices and in whether the biggest customers keep raising money. History offers both endings.

What would change the picture

The signals to watch are about credit and returns, not just rates.

• CoreWeave refinancing at sensible rates, or its bonds recovering toward par, would show lenders are still willing; Nebius’s Citi loan on Oct 22 is an early test.
• A large OpenAI funding round, or 2027 capex plans backed by cash flow rather than debt.
• A Fed pause and lower long-term yields.
• Warning signs: a sustained fall in GPU rental prices, a major GPU write-down, a cancelled contract or a covenant breach.

All 66 stocks in our group

These are the 66 stocks from our screen of US-traded stocks that fell 40% or more from their highs of the past six months and are tied to the AI build-out (most chipmakers excluded), with data from our Oct 23 screen, sorted by group and then by value lost; together they lost about $2.04 trillion (our math).

Moves are from each stock’s highest close since April 23 to the Oct 22 close. Seven foreign companies use their home listings (Tokyo, Paris, Madrid) and local-currency moves, with value lost converted at dollar prices; on that basis ACS is down 38%, just under the screen’s 40% cut, which its US ADR met. SoftBank and Murata use their US ADRs, as in the table above. Value lost is current market value scaled back to the peak; rows are rounded, so they may not add exactly to the totals.