AI Fundamentals Have Lost a Leg: Phase Two Confirmed, Phase Three Begins

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.

Two weeks ago I reduced what the market is pricing to a single question: where does the money to build data centers come from, and will the cost of borrowing it rise high enough to make the projects uneconomic. I split this correction into four phases. Phase one is position unwinding. Phase two is the repricing of financing costs. Phase three is capex actually slowing while revenue fails to catch up. Phase four is a re-rating of the business model itself.

Last week’s conclusion: layer two has entered its second half, layer three has produced its first evidence, and layer four has shown an early signal.

The numbers and their shape

The chart above covers last Friday, last week, the month of October, and Monday's session. I won't walk through every figure. All four panels say the same thing: AI-linked mega-cap tech fell, and everything else rose. Money did not leave the stock market. It moved out of AI and into the rest of it. One detail worth noting: the Dow lost 1.1% on Friday, the first time in two weeks that selling spread beyond AI. Defensives failed to catch the fall on Friday, then caught it again on Monday.

The financing window is closing one door at a time

Door one, high-yield and private credit: closed. OpenAI set out to raise $30 billion and collected roughly $11 billion, with its valuation cut from $1.4 trillion to $1.25 trillion. CoreWeave, Nebius, and the other neoclouds that borrow to build compute can no longer issue in the bond market. Blackstone told its investors that the $63 billion chip financing for Anthropic will likely carry a high-yield rating once listed; investment-grade bonds don’t come with high-yield coupons. Nvidia is in talks to provide a credit guarantee for a $45 billion data center that a neocloud is building for Anthropic.

What these headlines describe is one pattern: the customer can’t borrow, so the supplier steps in to guarantee, and the customer’s credit risk migrates onto the supplier’s balance sheet. Two weeks ago it was Broadcom backstopping Anthropic’s chip financing. Now it’s Nvidia guaranteeing a neocloud’s project. In the 2000 telecom bubble, Lucent lent its customers money to buy Lucent equipment; when the customers defaulted, Lucent went down with them. The legal form is different today. The economics are identical:

The seller is funding the buyer’s purchases. Part of the seller’s revenue is its own money taking a round trip.

Door two, equity issuance: the discount is widening. On October 26, four financings landed in a single day, and the market rallied for three sessions.

1. Meta sold $95 billion of stock.
2. SpaceX sold $40 billion of bonds to buy Nvidia chips.
3. Broadcom led a $63 billion package for Anthropic to buy chips.
4. Anthropic disclosed $77 billion in financing commitments.

Now look at it from the buyer’s side. Google’s June stock sale is trading 22% below the offering price. Intel’s August sale, 33% below. The first tranche of Meta’s Monday offering was underwater by more than 2% within a week.

When a company sells new shares, the buyers are hedge funds, long-only managers, sovereign funds. They buy at the offering price expecting the stock to trade above it. All three large offerings lost money for everyone who participated. The next time a company comes to market, buyers will demand a bigger discount before they commit. The bigger the discount, the less the company raises per share and the more existing holders get diluted. The cost of equity is rising, and it’s rising one loss at a time, written into the buyers’ books.

Credit issuance in 2026 is running at three times 2025. Supply has tripled, and the buyers are all losing money.

Door three, investment-grade bonds: still open, and it won’t close for several quarters. Google sold €6.7 billion of bonds on Monday with orders above €17 billion. Meta plans at least $20 billion in dollar bonds. Microsoft plans $25 billion, its first issue since 2017. The buyers of investment-grade AI debt are insurers and pension funds. They’re buying duration and a rating, not an AI story; as long as the rating holds, the demand holds. But all of them are borrowing, and that tells you companies with positive free cash flow also need outside money to keep pace with capex.

Last Wednesday Oracle told its investor day that it would raise only $20 billion more over the next four years. The stock rose 10.6%, the only cloud major rewarded in two weeks. Four trading days later it sold $10 billion of stock and $10 billion of bonds in one go. The stock fell 8.9%.

From Wednesday to Friday the market’s standard tightened three times: negative free cash flow, sold; positive free cash flow but issuing debt, sold; capex up while revenue down, sold. Monday added a fourth: promised not to borrow, then borrowed, sold hardest.

The market is now asking one question: did you try to borrow money this week?

Marvell buying Lumentum with stock is an early layer-four signal. A $115 billion all-stock deal at a 5% premium. Lumentum is up 900% in a year and agreed to sell at essentially no premium. Marvell agreed to dilute itself by a third to buy a company at a cyclical peak. Neither side haggled; both signed. That behavior is only rational if both believe their own stock is expensive. When companies in a sector start using their shares as currency, that’s layer four beginning.

For readers new to the framework, the four layers come from: When the Music Stops II (Street Alpha, Oct 25).

AI’s cost is starting to squeeze the companies that pay for it

Intel’s problem is more complicated than weak CPU demand. The quarter itself was good. Data center revenue grew 78%, the fastest in 15 years; AI demand for server chips is intact. What was bad was the fourth-quarter guide: revenue below the third quarter, gross margin down sharply, and 2027 capex rising significantly with free cash flow turning negative.

Global PC shipments fell about 20% year over year in the third quarter. Intel’s CFO called it the worst third quarter he’d seen in 20 years. Two things are going on. AI data centers have bought up the memory supply; memory prices have pushed up the cost of a finished PC; corporate IT budgets are fixed, so demand gets crushed. And the PC makers spent several quarters building inventory and are now clearing it, so Intel’s shipments are falling faster than end consumption. Management says the destocking is heaviest in the fourth and first quarters.

So Intel is a layer-three signal, but the precise reading matters: it is not that AI demand is weak. It is that AI’s demand is destroying the economics of a different industry, and Intel gets half its revenue from that industry. Dell, HP, HPE, and NetApp all fell together on Friday; the common factor is memory cost. There’s also a link back to layer two: Intel sold $23 billion of stock at roughly $100 in August. The stock is under $60.

Apple is the biggest sample on the memory line. After the close on Monday, Apple beat on revenue and earnings, with iPhone up double digits for a fourth straight quarter. The stock fell 3%. The reason was the guide: December-quarter gross margin more than a point below consensus, and services growth slowing. The new CEO’s exact words: “the memory pricing we’re seeing is unprecedented,” with costs rising beyond the December quarter. The company also called out higher inference costs tied to Siri’s AI. Apple has no shortage of cash and no need to raise any. Its decline is not layer two. It’s layer three: AI’s cost compressing the margins of a company that isn’t building data centers.

Anthropic’s amended S-1. Three things happened at once. Sequential revenue growth slowed from 144% last quarter to 47%. Compute cost rose from 48% of revenue to 58%, and adjusted operating margin fell from 19% to 15%. Total compute commitments rose from roughly $560 billion a week ago to $895 billion, of which $437 billion is non-cancellable and roughly $213 billion carries termination rights.

Anthropic buys compute from cloud providers. For the first two quarters, Amazon and SpaceX charged discounted rates; in the third quarter the discounts expired and billing moved to market rates. On top of that, about $5 billion of third-quarter compute was bought on short-term contracts, which the company itself says are far more expensive than multi-year deals. Contract payments are fixed. Revenue growth is slowing. The gap between the two is widening. More contracts start billing in the fourth quarter, and the company disclosed that usage drops roughly 20% seasonally in the last two weeks of the year.

This is the healthiest AI lab financially: positive operating cash flow and $130 billion in cash. Its non-cancellable commitments equal 26 quarters of revenue. The numbers say there is no room for error in 2028. For weaker labs and for neoclouds carrying real debt, the room runs out in 2027. Anthropic’s own risk factors say that some of its compute suppliers depend on third-party financing or credit support to buy chips and build capacity, and that if that financing becomes unavailable or more expensive, contracted capacity could be delayed, repriced, or not delivered. That sentence is the company itself connecting layer two to layer three.

AI is eroding the businesses that pay for it. Google’s search growth slowed from 17% to 12%; one reason the company gave is that ads in AI answers monetize worse than traditional search. Reddit fell 29% on lower referral traffic from Google search. Intel’s and Apple’s hardware margins are being squeezed by memory prices. The common thread: AI is eating the revenue and margins of other businesses, and those businesses are precisely the cash flows that fund AI capex.

The fundamentals have lost a leg. S&P 500 earnings grew 30% in the third quarter, and 2027 estimates are still being revised up. What has broken is the free cash flow of the specific group of companies paying for AI capex. The five hyperscalers’ combined 2027 capex will exceed their free cash flow; for every additional dollar of operating cash flow, they spend $1.57. In the third quarter, four of them already generated less operating cash flow than they spent on capex. A large share of the S&P’s earnings growth comes from the sellers, Nvidia and TSMC, and from fair-value gains on the AI stakes held by Google and Amazon.

Capex keeps accelerating. Meta, Google, and Amazon all say 2027 will be significantly higher. The ability to pay is declining. The market sees the gap widening and has not seen a single company name the date it starts to close.

Correlation among the Magnificent Seven has fallen to a ten-year low. Microsoft and Google reported the same night with cloud growth in the same range; one rose, one fell. The common factor, AI demand, no longer decides direction. Each company’s cash flow coverage and financing structure does.

Funding markets and market structure

For readers who follow macro, two rates to start with.

What the Fed pays banks on their deposits is called IORB, currently 3.90%. A bank leaves cash at the Fed, does nothing, and earns 3.90% a year.

SOFR is a different rate: someone pledges Treasuries as collateral, borrows cash overnight, and repays the next morning. The borrowers are mostly hedge funds and large dealers. The lenders are mostly money market funds, which hold trillions in cash but aren’t allowed to deposit at the Fed, so they have to lend it to someone.

Normally SOFR sits below IORB, because money funds have nowhere else to put cash and will lend it a little cheaper. Banks borrow that cheap cash, park it at the Fed for 3.90%, and pocket the spread. As long as banks are willing to run that trade, SOFR stays pinned below 3.90%.

Last week SOFR closed above 3.90% for five straight days. That’s when you start asking whether the system is developing a liquidity problem.