The Nasdaq 100 just posted its worst week in months. Not because AI demand disappeared. Because the debt required to deliver that demand suddenly costs 15 percent.
CoreWeave’s Bonds Named the Fracture First
On October 16, Reuters reported that CoreWeave bonds issued in June at a 9.675 percent yield had fallen far enough in price to push the yield near 15 percent. Five-year credit-default swaps on the name climbed past 1,000 basis points and reached as high as roughly 1,200. Those levels imply a market-priced probability of default near 60 percent inside five years.
The company carries approximately $50 billion in debt and lease obligations. Shares closed the day down 11.91 percent at $61.39, capping a week in which the stock lost roughly a quarter of its value.
Management has pointed to a contracted backlog exceeding $100 billion, much of it with investment-grade counterparties, plus an existing Nvidia arrangement covering unsold capacity through 2032. The credit market still demanded distressed yields. That is the tell.
Capex and interest expense arrive now. Revenue arrives later. When the cost of bridging that gap jumps from high-single-digit to mid-teens, the entire financing model for the neocloud layer gets re-examined.
For how this credit stress first hit chip stocks earlier today, see our note: AI Debt Concerns Just Stress-Tested the Compute Trade.
Semiconductors Absorbed the Longest Slide Since 2008
Chip stocks fell for an eighth consecutive session on October 16, the longest losing streak since October 2008. The PHLX Semiconductor Index closed at 11,137.29, down 3.83 percent on the day.
From its last positive close on October 6 the index had declined roughly 16 percent. The VanEck Semiconductor ETF (SMH) finished the week down about 10 percent, its worst weekly stretch since April 2025.
Micron, AMD, Intel, and Broadcom all declined between roughly 4 and 6 percent on Friday. Memory names took additional pressure from reports of progress by China’s CXMT on advanced DRAM architecture.
Demand for HBM and advanced logic has not vanished. What changed is the multiple investors are willing to pay while the debt that funds the data-center buildout is being re-priced in real time.
For the broader financing picture behind the neocloud layer, see The AI Buildout Hits Its Funding Wall.
Rates and Oil Kept the Pressure Elevated
Hotter-than-expected inflation data earlier in the week did not help. September CPI printed 3.7 percent year-over-year. The PPI rose 1.3 percent month-over-month, the largest monthly gain since 2022.
Federal Reserve Chair Kevin Warsh, speaking at the IMF meetings in Bangkok on October 16, declined to pre-judge the late-October or December meetings while reiterating the priority of price stability. He described the AI investment surge as a near-term demand shock that is complicating the path back to target. CME pricing placed the probability of a December rate hike in the mid-70s percent range after the data.
Oil prices jumped on the same day. WTI settled near $101 and Brent near $110 on geopolitical supply concerns.
Higher energy costs feed directly into data-center operating expenses and into the broader inflation readings the Fed is watching. Rate-sensitive growth names, already carrying elevated leverage, felt the combination first.
The Nasdaq 100 itself closed October 16 at 29,187.55, down 2.50 percent from the prior session’s 29,935.59. For the week the index lost approximately 5.5 percent, its worst five-day stretch since June. The broader Nasdaq Composite fell 2.15 percent on the day and roughly 4.5 percent for the week.
Optical and Power Constraints Have Not Disappeared
Financing is the near-term fracture. The physical stack is still tight. At SEMICON West on October 13, Nvidia principal researcher Nikola Nedovic presented a prototype of optics-on-interposer technology that places optical engines directly on the interposer carrying the compute dies, eliminating the organic substrate path.
The demonstration used eight wavelengths at 32 Gb/s each, delivering 256 Gb/s per fiber. The goal is higher bandwidth density at lower power. It remains research-stage.
Separately, Meta disclosed at OCP that its 51.2T co-packaged optics evaluation had accumulated 1.8 million system-hours, with link interruptions running three times lower than pluggables in the first phase and six times lower in the second. Roughly 85 percent of remaining interruptions traced to a single external laser-module fault that Meta said had been addressed. These are early deployment data points, not commercial scale, but they illustrate the industry is still trying to solve density and power problems that debt alone cannot fix.
On the power-semiconductor side, Infineon and Wolfspeed signed a multi-year agreement around October 15 for 200-millimeter silicon-carbide wafers. The deal supports higher-voltage, higher-efficiency power devices used in AI data-center power supplies and other high-performance applications. SiC and GaN continue to see structural demand from the same infrastructure buildout that is now facing higher financing costs.
The Open Question
The credit market has put a number on the timing mismatch. Equities have only started to react.
Contracted backlog is real and large. Whether that backlog can be financed and delivered at a cost of capital the market is only now accepting, or whether the buildout simply slows as refinancing and new debt become more selective, is the question the next few quarters of capacity deliveries and credit spreads will answer.
The demand signals have not collapsed. The financing math just got harder. That distinction is what the tape is still sorting.


Leave a Reply