The selling began in servers and reached the chips within a day.
Before U.S. markets opened on October 15, Nvidia, Advanced Micro Devices and Micron Technology were all quoted lower, and Nasdaq futures had slipped with them. The pressure traced back to the prior session, when Dell Technologies closed down 11 percent and pulled Hewlett Packard Enterprise, Oracle and Super Micro Computer down alongside it. A question about one hardware maker’s order book had become a question about who ultimately pays for the buildout.
Barclays supplied half of the answer. In a note cutting its ratings on Dell, HPE and NetApp, the bank cited two things: the scale of the rally that preceded the call, and the timing of enterprise adoption of artificial intelligence. Dell’s shares had climbed more than 300 percent at one point this year. HPE was up roughly 160 percent. Both ranked among the largest gainers of the data center cycle, which is another way of saying both were priced for a spending curve that has not yet arrived.
The other half of the answer sits in the credit markets. Broadcom has been arranging large debt packages for customers financing custom chips and the buildings that hold them, according to Crypto Briefing. A package for Anthropic totals about $60 billion, split into $42 billion of senior secured debt and $18 billion of junior debt led by Blackstone. Discussions with OpenAI cover $30 billion to $50 billion and remain at an early stage. Together the two commitments could exceed $92 billion. Broadcom established an AI financing platform with Apollo and Blackstone in June, and the first $35 billion transaction went to Anthropic.
Investors spent the morning repricing what that borrowing implies. CoreWeave, the most heavily indebted name in the group and the one with the least room to wait, closed down 15.5 percent on October 14 after falling as much as 16 percent during the session. Its unsecured notes maturing in 2032 traded below 80 cents on the dollar, according to Trace, pushing the yield above 15 percent. Credit default swaps on the company reached a record. Market pricing implied a better-than-even chance of default within five years, and one model cited in the market put the figure near 63 percent.
Mike Intrator, CoreWeave’s chief executive, told employees in an internal note that technology companies of a new type always pass through a stretch of market misunderstanding and irrationality, and that the company’s contracted backlog covers its debt in full. S&P rates the company B plus. Gil Luria of D.A. Davidson described it as the subprime tranche of the AI trade and warned that the shares could fall to zero if the financing wave recedes.
What makes the selling awkward is that the demand signals published the same week point the other way. TSMC reported third-quarter net profit of 763.1 billion New Taiwan dollars, about $24 billion, up 68.7 percent from a year earlier and its sixth consecutive record. Revenue rose 51 percent to 1.49 trillion New Taiwan dollars. Operating margin reached 58.7 percent, above the top of the company’s own guidance range. TSMC raised its fourth-quarter revenue guidance to $52.3 billion to $54.3 billion, a midpoint that implies 14 percent growth from the prior quarter and 58 percent from a year earlier.
The argument in the market, then, is not that chips are going unbuilt. It is that the companies ordering them have not converted the spending into revenue yet, and that the difference is being carried on borrowed money at a moment when lenders are asking for more.
Labor data offer one measure of the gap. Lightcast counted more than 7,300 U.S. job postings for forward-deployed engineers in the first nine months of this year, against fewer than 200 in all of 2023, according to CNBC. The median advertised salary was $195,456, more than three times the national median across all occupations. AWS has committed $1 billion to the function. Anthropic has set aside $100 million to train 10,000 forward-deployed engineers by the end of 2027. OpenAI’s deployment business opened with more than $4 billion of investment. Deloitte carried the most openings with 872, followed by Accenture at 688 and Google at 499.
Those roles exist because a model that answers a question is not the same product as a system that closes a purchase order, files a claim or reconciles a ledger. The distance between the two is where the payback question lives, and it is the distance Barclays pointed at.
Money did not leave the sector on October 14 so much as move inside it. Apple closed up 0.5 percent at $342.96, giving it a market value of $5.01 trillion and making it the second company after Nvidia to hold that level through a closing bell. Oracle fell 8.83 percent, Meta 3.27 percent, Nvidia 3.16 percent and Microsoft 1.8 percent. The gainers were the companies whose revenue comes from buyers paying for products rather than from buyers financed to purchase them.
Whether that distinction survives the next round of earnings is unresolved. Dell reports an order book that depends on corporate technology departments showing up. Broadcom reports a chip business whose largest customers borrow to place their orders. CoreWeave reports into a credit market that has already voted.
None of the three has yet had to answer whether debt raised against AI capacity gets repaid out of that capacity’s revenue. October 15 was the day investors began pricing the possibility that it does not.
