The thread
In the six trading sessions to October 15, the Philadelphia Semiconductor Index fell about 12%, roughly $1 trillion of market value. I went through the 189 highest-voted comments in a WallStreetBets thread about the selloff and sorted them by the case each one made. WallStreetBets is the largest retail trading forum on Reddit, so this is a sample of the most active retail traders rather than the whole market, but it shows how they were reasoning in their own words.
About a third of the comments were bullish and buying the dip. Seven percent said the bubble had popped. Roughly half blamed outside events, mostly Iran and the midterm elections, or hadn’t made up their minds. Only 8.8% laid out the financing argument this post is about, in which higher borrowing costs make projects harder to fund, delays follow, the weakest developers fail, and chip demand softens later. Even among the bears, it was a minority view.
The flows pointed the same way. Retail traders on Citadel Securities’ platform bought a net $2.6 billion of semiconductor stocks in the week ending October 9, their third-largest weekly total in a year. The following week, investors put $6.1 billion into SOXL, a fund that returns three times the daily move in chip stocks, more than went into any other US-listed ETF, according to ETF.com. Over the same week the unleveraged chip fund SOXX lost $1.2 billion to redemptions and fell 11.8%.
The bulls’ case came down to one challenge: show me actual cancelled orders.
It’s a fair challenge, and the bulls had their facts right. On October 15, TSMC reported record third-quarter profit, guided fourth-quarter revenue to between $52.3 billion and $54.3 billion, and its CEO said AI demand was still well above what the company could supply. Nobody had cancelled anything.
The problem is the question those facts were answering. The bulls were checking whether demand for AI is still there, and it is. What moves prices in a build-out paid for with borrowed money is whether the buyers can keep paying. For most of a cycle those two questions have the same answer. They come apart when borrowing gets expensive, and cancelled orders are the last place that shows up.
When spending passed cash flow
Stijn Van Nieuwerburgh’s paper for the fall Brookings Papers on Economic Activity puts a date on the shift. Oracle, Microsoft, Amazon, Meta and Alphabet spent about $97 billion on capital projects in 2020 and more than $400 billion in 2025, and they’re projected to spend more than $800 billion in 2026. On current estimates, 2026 is the first year that spending exceeds their combined operating cash flow. From here, part of the build-out has to be paid for with someone else’s money.
Relative to the size of the economy, the paper finds this build-out larger than the canal, railroad, electrification, highway and telecom booms. Its central scenario has AI infrastructure investment reaching $10.3 trillion from 2025 through 2032, an average of 3.63% of US GDP a year, against 2.24% for the railroads and 1.10% for telecom and fiber. Railroads lasted for decades. The paper notes that GPUs may become economically obsolete within three to six years, which gives a lender far less time to wait for a project to pay off.
To earn a 10% return at a 50% operating margin, the build-out needs about $3.7 trillion of annual revenue by 2032, roughly 9.2% of GDP. If the equipment lasts three years instead of six, that rises to about $6 trillion. The paper benchmarks the requirement against about $100 billion of combined OpenAI and Anthropic revenue this year, which would have to grow about 80% a year to get there.
Who is paying for it
The paper cites a Morgan Stanley estimate that more than half of the roughly $2.9 trillion of computing capacity needed from 2025 through 2028 will be paid for with outside capital, split about 60% equity and 40% debt. Private credit is the largest lender, at about $800 billion, against about $200 billion of corporate bonds and $150 billion of structured finance.
There’s also a large amount the balance sheets don’t show. Moody’s estimates the hyperscalers have about $970 billion of lease commitments, and about $660 billion of that isn’t on their balance sheets yet.
Meta’s Hyperion campus in Louisiana shows how these deals are put together. The project is about 2 gigawatts and $30 billion of assets. Meta sold 80% of it to Blue Owl for about $2.5 billion, and the project company raised $27 billion of A+ rated debt, about 90% of the asset value, at a 6.58% yield. That’s at least a percentage point more than Meta would likely pay to borrow on its own, which the paper estimates at more than $5 billion of extra interest. Meta is paying that premium to keep the debt off its own balance sheet, and the paper suggests the deal may become a template for future data center financing. The project’s cash flow covers its debt payments 1.12 times, which leaves little room if the data center earns less than planned.
The lenders behind these structures are already rationing their own cash. Blue Owl has been limiting withdrawals from its private credit funds to 5% a quarter, against redemption requests of 17% to 39%.
What hedge funds and lenders were doing
In the five sessions through October 15, hedge funds trading through Goldman Sachs’ prime brokerage cut technology stocks at the fastest pace since Goldman’s records began in 2016, and semiconductors took the largest dollar hit. Even after the selling, tech made up 23.5% of Goldman’s global prime book, in the 90th percentile of the past five years, and net leverage fell to 48.6%, a one-year low. The funds were cutting risk in a crowded trade, and they still held a lot of it.
Credit markets moved more sharply. In an October 23 note using Bloomberg data through October 21, J.P. Morgan Private Bank showed five-year default protection on the North American technology, media and telecom index above 100 basis points, against about 60 for the broad investment-grade index. That gap of roughly 40 points is wider than it got during the 2022 tech downturn.
The repricing was uneven. Default protection on large AI companies that generate plenty of cash traded between 50 and 120 basis points. Oracle’s was near 300. CoreWeave’s passed 1,000 in mid-October and reached 1,100 the following week, a level J.P. Morgan reads as roughly even odds of default over five years, and its 2032 bonds traded near 78 cents on the dollar. Lenders were separating the companies that pay for AI out of their own cash flow from the ones that have to keep borrowing to do it. The banks in the middle weren’t spared either. Goldman Sachs, whose prime book is the source of the hedge fund data above, saw its own shares fall 20% even as it reported record earnings.
The data can’t show that retail was buying what hedge funds sold, because it doesn’t say who traded with whom. What it shows is two groups looking at the same weeks and drawing opposite conclusions.
Why the funding side breaks first
In a build-out financed with debt, the trouble tends to show up in a set order. Lenders start charging more. Projects that already have their funding keep going, so orders stay full and earnings stay strong. New projects stop making sense at the higher cost, so fewer get started, loan syndications shrink, and listings get pulled. Suppliers cut their guidance after that, and cancelled orders come last.
That’s why the dip-buying in the thread can keep going for a while. Demand is the last thing to change, so the demand news stays good well after lenders have started pulling back. Waiting for cancelled orders means waiting for the final step, and by then prices have usually moved.
The closest precedent is 2007. Indices tracking subprime mortgage bonds started falling early that year and collapsed over the summer. In March, Fed Chair Ben Bernanke told Congress that “the impact on the broader economy and financial markets of the problems in the subprime market seems likely to be contained.” The S&P 500 didn’t peak until October 9.
How a slowdown would land
A funding squeeze wouldn’t hit every company in the AI build-out the same way.
| Company | What a funding squeeze does to it |
|---|---|
| Chipmaker | Can stay profitable while expectations for its growth fall |
| Hyperscaler | Can absorb a bad bet out of its own cash flow |
| Leveraged data center developer | Can fail while still owning useful assets |
| AI application company | Can grow usage without growing its margins |
What would prove this wrong
The clearest sign this is wrong would be new AI project financings closing on terms no worse than this summer’s. The gap between tech and broad investment-grade default protection narrowing back toward its 2022 range would count against it too. So would the hyperscalers raising their 2027 spending plans and paying for the increase out of their own cash flow.
If it’s right, the first signs will come from the funding side, such as a data center listing that gets pulled or a private credit fund selling AI loans below face value. A hyperscaler project financing that can’t find enough buyers would be the strongest signal of all.
Sources
- Stijn Van Nieuwerburgh, “Financing the AI Buildout,” Brookings Papers on Economic Activity conference draft, September 2026
- J.P. Morgan Private Bank, “Credit markets look for AI’s weak links,” October 23, 2026
- Reuters, “AI credit fears push tech default insurance to record highs,” October 15, 2026
- ETF.com, “Investors Buy the Dip in Chips,” October 19, 2026
- Citadel Securities retail flow data, week ending October 9, 2026
- Bloomberg, “A $1 Trillion Chip Rout Draws Billions Into Leveraged ETF Bets,” October 15, 2026
- 24/7 Wall St., “Hedge Funds Dumped Tech Stocks at the Fastest Pace in a Decade, Goldman Says,” October 20, 2026
- CRN Asia, “TSMC third quarter revenue tops guidance as CEO says AI demand still exceeds supply,” October 2026
- Bank for International Settlements, “The pricing of subprime mortgage risk in good times and bad: evidence from the ABX.HE indices”
- Benzinga, “Ben Bernanke Tells Congress Housing Crisis Is Likely ‘Contained’”