When the Music Stops II: Near-Term Outlook for US Equities and an update to the Exit Framework for AI

Disclaimer: This article is for informational and educational purposes only. It does not constitute financial, legal, or investment advice. Views are based on public information as of the publication date and are subject to change.

Over the past two weeks, AI-related stocks suffered their largest drawdown since April 2025, while the rest of the market barely moved.

• The weekly headline numbers look mild: S&P 500 −0.5%, Nasdaq −0.7%, Dow +0.4%.
• They hide Thursday’s massacre: Nasdaq −2.6%, the semiconductor index −6.72%, its worst single day since June.
• Over two weeks the Nasdaq is down 5.2% and the SOX roughly 14%, leaving it 27% below its June high. The S&P sits just 2.5% below its October 6 all-time high.

The market is now separating two things it used to treat as one: AI demand and AI financing. What we are living through is the repricing of the second.

Many readers still seem to miss this, so here is the simple version: At first, everyone cared about how fast AI was growing. Then AI started borrowing. Then interest rates spiked, and the market began asking a different set of questions:

Where does the money come from? Will borrowing costs rise to the point where projects don’t generate enough returns? Can the borrowers actually deliver the projects, and repay on time?

That worry has been climbing the credit ladder, one rung at a time, from the weakest borrowers to the strongest:

1. Last week, CoreWeave. The most levered compute landlord. Its bonds fell to 78 cents on the dollar; CDS blew past 1,000 basis points.
2. Monday, Nscale. IPO valuation cut from $30 billion to $16–19 billion.
3. Wednesday, SpaceX. Disclosed next-quarter capex of $35–40 billion against roughly $51 billion of cash. CDS widened.
4. Thursday, Meta. Disclosed $825 billion of contracted future spending, stated in a filing that it may need to raise “hundreds of billions,” and saw its CDS hit 115 basis points, the widest on record.

Thursday also brought the Bloom Energy report everyone was watching. Record profits, but operating cash flow of negative $27 million, free cash flow of negative $180 million, and a 19% drop in two sessions.

What Bloom’s quarter exposed is vendor financing: Bloom is fronting money for its customers. Of $1.07 billion in product revenue, $430 million was equipment already shipped and already booked as revenue, but not yet paid for, because the contracts say payment comes after installation or energization. On the balance sheet this is called a “contract asset,” and it doubled in a single quarter.

Why ship before getting paid? Because Bloom deploys faster than its customers can absorb. Its pitch is power in months, so the equipment has to be on site early, while the customer’s grid connection, permits, and building work lag behind.

But that is not the real question. The real question is why customers are slow? Is it power and permitting, or is it that the customers themselves are short of financing and slowing their projects? Bloom’s largest customer is 43% of revenue, very likely Oracle, and Oracle’s bonds traded down to 86 cents this month. There is no answer yet. The market has priced the worst case first.

The point I want to make is this: the market is now acutely sensitive to every financing question around AI capex. If a company’s earnings report does not directly answer “where does the money come from, and how do you pay it back,” the report becomes a sell-the-news exit.

This is the repricing of AI financing I described above. It will take months, and with incomplete information, institutional investors will stay on the sidelines.

Weekly tape (skippable): Monday through Wednesday was a bounce, Thursday erased all of it, Friday stabilized without recovering.

• Monday (10/19): A relief bounce with no real buying. S&P +0.41%, Nasdaq +1.11%. Evidence: Goldman upgraded CoreWeave to Buy and JPMorgan defended its debt the same day; the stock still closed lower at a new low. Good news that doesn’t lift a stock means sellers are still there.
• Tuesday (10/20): Rotation day. AI rested, old economy rallied. Dow +1.02%, Russell +1.65%, Nasdaq flat. After the close, SpaceX and Super Micro delivered the first firsthand demand-side evidence in two weeks: SpaceX compute revenue up $5 billion in a quarter, Super Micro raised revenue guidance by $4 billion.
• Wednesday (10/21): The only day in two weeks that AI stocks rose on good news. S&P +0.58%, SOX +1.97%, SpaceX +7.8%, Super Micro +9.6%. Then three after-hours reports pushed the conversation from “demand” back to “borrow and repay”: Meta raised the floor of its capex guide and disclosed $825 billion of commitments; Tesla cut capex by $5 billion and pushed its robot timeline out a year; Lam Research said component shortages are limiting shipments and guided only in line.
• Thursday (10/22): The semiconductor massacre, the day the market repriced “borrowing.” Nasdaq −2.6%, SOX −6.72%, Lam −11.8%, Meta −7%, Tesla −3.1%, CoreWeave −10%. Three reports hit three different nerves: Tesla said demand may be delayed, Lam said supply isn’t accelerating, Meta said the biggest buyer is about to borrow. Yet the Dow was flat, the equal-weight S&P rose, and industrials, rails, defense, and staples all gained. Only AI fell. Gold, silver, bitcoin, and the Nasdaq fell together, which signals forced liquidation across all positions, not a flight to safety. After the close, Western Digital beat across the board, and the CEO said incremental demand is coming from neoclouds and AI labs. In other words, from borrowers.
• Friday (10/23): AI stabilized, oil took over. SOX +0.44%, with an intraday low almost identical to Thursday’s; both held at the 200-day moving average. WTI +3.8% to $103.27, Dow −1.23%, and Thursday’s winners became Friday’s losers. The French 10-year hit 5.13%, with the France-Germany spread at 175–178 basis points, the widest since 2011. OpenAI cut the valuation target for its funding round from $1.4 trillion to $1.25 trillion.

The Market Is Worried About Two Opposite Things

Half the market fears compute cannot be delivered. The other half fears too much will be. What the two fears share is a timing gap: capex is spent now, revenue arrives only after the power is connected and the customers show up. In between, you either carry the financing cost with no revenue, or the revenue arrives after prices have already fallen. The higher the interest rate, the more expensive that gap becomes.

The under-delivery fear is about power. Estimates for new data center capacity next year range from 20 gigawatts to 40–50 gigawatts, a gap of more than double. Lam Research says shipments are constrained. Oracle sent force majeure notices to its financing partners because it couldn’t get grid connections. If chips are built but data centers can’t be energized, buyers who prepaid for chips get no revenue, and equipment makers’ order recognition slips.

The over-delivery fear is about price. If 40–50 gigawatts really come online, compute supply doubles and rental and memory prices fall. Rents for the newest Blackwell generation are still making new highs, but on-demand H100 pricing has fallen from $8 per hour in the first half of 2024 to about $2.64. New generation up, old generation down: this is what the early stage of rising supply looks like.

Right now, prices are being set by rates and positioning, not fundamentals. Western Digital beat on every line Thursday night, and the CEO’s remark that new demand comes from neoclouds and AI labs sent the stock down, because those words now mean “customers who borrow.” Fundamentals are not the marginal driver of price today. I wrote about this dynamic in August; if you want the longer version.

What Comes Next

Here’s my framework. The past two weeks of selling can be separated into four layers, each on a different time scale.

Layer one is position liquidation, measured in days to weeks.

Hedge funds had their worst week since 2020; Goldman’s crowded-positions index fell 10% in October. This layer is nearly done, based on dealer-desk data: JPMorgan notes hedge fund gross exposure has already come down sharply, leaving less room for concentrated selling.

Layer two is the repricing of funding costs, measured in weeks to months.

This week’s events, from CoreWeave to Meta, are this layer. It is underway.

Layer three is an actual slowdown in capex, measured in months to quarters.

It hasn’t started. Every data point says AI capex is still accelerating. What the market is pricing is that it cannot continue at today’s cost of capital. Tesla’s $5 billion cut was the first, but Tesla is not a hyperscaler.

Layer four is the rerating of the business model, measured in years.

For twenty years, Microsoft, Google, and Meta were worth what they were because they could grow without spending: software written once sells infinitely, and a network gets harder to challenge the bigger it gets. Investors paid up for two things, no capex and no competitors.

AI takes both away. These companies now spend over $100 billion a year building data centers, turning them into capital-intensive businesses in the mold of 1840s railroads or 2000-era telecoms. And within a year of AI’s arrival, three or four roughly equal frontier models exist. No one has a monopoly.

The last time this happened was telecom around 2000.

Internet traffic really was doubling every year; demand was real. Carriers borrowed hundreds of billions to lay fiber, and equipment vendors lent customers money to buy their own gear. When credit tightened, the leveraged carriers couldn’t service their debt. More than half of North American high-yield defaults in 2001–2002 came from telecom, and bondholders recovered about twenty cents on their dollar. The fiber wasn’t wasted; streaming and cloud filled it a decade later. But the people who funded it didn’t get their money back. Vodafone was Britain’s largest company by market cap in 2000 and still trades below that peak today.

I’m not telling this story to frighten anyone. The internet didn’t fail. The fiber didn’t fail. But the lesson is this:

When an industry goes from needing no capital to needing a lot of borrowed capital, and from having no competitors to having many, the multiple the market pays for it falls, and it does not wait for earnings to decline before repricing.

This week, Meta beat on revenue with accelerating ad growth and fell 7% because it may need to issue equity. That is the first signal.

If you are reading carefully, please build your own sense of risk. Watch for events like this and remember them. They are the first-stage signals in our exit framework.

My Predictions for the Next Two Weeks

1. A bounce will come, but it will be limited and it will be sold.