NVDA shakeout playbook: memory, cash, Berlin

Watching Friday's close, I saw once again how fast selling spreads across AI stocks. NVDA slid from a $243 peak to the $215 area, down about 11 percent in ten sessions. Investors question whether infrastructure outlays will ever pay back while yields climb and risk appetite fades. I watch this tape with curiosity rather than fear, because the whole group got hit for the same stated reason. Through the whole recording I kept asking one question: did fundamentals break, or did pricing overshoot. My answer leans to the second, though that is no green light for blind buying.

CNBC reports CoreWeave's 2031 and 2032 bonds now yield close to 15 percent, and the shares lost more than 15 percent in a single day. The distinction matters: coupons stay fixed at 9.75 and 9.625 percent, so cash leaving the company has not changed; what changed is the cost of the next borrowing round. Selling spread to Oracle and SoftBank paper, and fear now covers the whole neocloud space. I look at balance sheets instead: NVDA and TSMC sit on huge cash piles with no debt strain. Leveraged growth stories are a different animal, and caution fits them better.

The DRAM corner looks like the likely winner of this mess. Micron fell hard, then started to recover; Barron's flagged the dip as buyable. Musk saying he would buy all memory output for three years shows the scale of demand. My read is simple: supply cannot keep up, and fresh capacity gets absorbed the moment it comes online. Talk of a memory glut remains unverified. I stay selective here, preferring large names with real revenue over adventures.

Wall Street analysis compiled by ChosunBiz suggests the CXMT 4F-squared push got overhyped. A Rosenblatt analyst calculates the Chinese maker's cell area could stay 34 percent larger than rivals' current 6F-squared cells, while yield and cost-per-bit figures remain undisclosed. Samsung and SK already demonstrated the same design yet held back from volume output over yield worries. Scaling without EUV tools adds another question mark. A product announcement is not the same as efficient volume production; markets missed that gap.

Three stories from last week were all misframed, in my view. First, OpenAI's $68B gross and $50B net revenue figures got mixed up, though nothing new had emerged. Second, the Broadcom chief's swipe at traders chasing frivolous headlines got presented as an insult to all AI investors. Third came the CXMT exaggeration. Yahoo's Anthropic numbers tell a similar tale: third-quarter revenue hit $16.8B, up 46 percent on the quarter, yet down from the prior 143 percent pace. Growth is still more than 3.5 times in two quarters; markets are pricing the change in speed, not the level.

The most interesting idea in the recording, to me, was the central-bank metaphor: Huang uses balance-sheet muscle to back startups and put his name on leases, cutting financing costs. Some call it circular financing, others read it as high conviction. Both sides have a point; the startup book looks measured next to the balance sheet. The signal I watch is clear: bank appetite for this credit at Goldman and Morgan. If they step back, the story changes; until then it stays noise.

The Wall Street bull camp is not backing down. Evercore ISI keeps its buy call, arguing Blackwell delay and demand fears are overdone. Yahoo reports Ives holding his $300 target, saying NVDA's moat now stretches past silicon into networking, software and next-generation systems. I think the core of that view is right; competition now runs across the whole platform, not one product. Still, price targets deserve to be read as scenarios, not promises.

Concrete news arrived from the PC side: RTX Spark notebooks are on sale. Asus, Dell, HP, Lenovo, Microsoft and MSI all list models starting at $2,599. A single processor pairing delivers 1 petaflop of FP4 compute with a 128GB shared pool, so large jobs and local helpers run without a constant cloud link. Photoshop and Premiere gained acceleration on the Adobe side, with DLSS 5 for games. My take: NVDA is moving past selling processors into selling the end unit itself, opening a new front against incumbents.

BloombergLaw first reported Taiwan's plan to build industry parks in the United States serving NVDA partners, starting in Texas at $2B to $4B per site. The Electrical and Electronic Manufacturers' Association takes the funding role while the state offers loan guarantees plus visa and land help. Giants such as TSMC and Foxconn can finance their own moves, but smaller suppliers cannot do it alone. I read this as part of the race to move the supply chain onto American soil in an age of cross-strait tension. Musk needling TSMC about chipmaking adds color.

Two different games got set up on the options desk. The bull side buys the November 6 225 call and sells the October 23 225 call, a calendar spread near $2.80 seeking a limited upside grind with a profit zone near $217 to $233. The bear side plays decline through an unbalanced October 30 put butterfly near $2.70. The technical picture: NVDA slipped behind its 50-day average near $223 and clings to the 100-day, with the 200-day near $203 as support below. I read those levels as a risk frame, not trade signals.

The macro tape is noisy: the VIX runs hot, oil pushes higher, yields sit elevated, and Fed moves plus the American presidential race all get priced. Friday's unwind pushed markets into support, and technician Newton sees the low landing any day now. The hardest question is whether the neoclouds survive this stress. Voices in the recording remind us CoreWeave's customers rank among the bluest chips, arguing the issue is funding cost, not demand. I buy that split, though high costs keep squeezing weak balance sheets.

Hard data keeps refuting the selloff story. ASML posted €11.7B in sales and €5.0B in net income for the third quarter, with order interest stretching into 2029 and 2030. TSMC revenue reached $46.72B, up 41.2 percent on the year, guiding $52.3 to $54.3B for the fourth quarter. B100 and B300 rental rates moved up in recent days, and DRAM prices held firm through the selling. Add the revenue surge at top labs, where capacity gets rented the moment it goes live. This reads as the inverse of the 2000s fiber excess: back then empty capacity waited for demand, while today demand waits for capacity.

The roadmap is clear: Blackwell Ultra has ramped, Rubin now reaches customers, Rubin Ultra follows in 2027 and Feynman in 2028. Huang talks of $3 to $4T in annual AI factory spending by decade's end, and the $3T revenue hint from the Fridman chat fits that frame. I believe NVDA can be worth far more in future years, but that is no guarantee and the path will stay bumpy. Eyes now turn to Berlin on October 21, where the opening keynote starts at 11:00 CEST and could add fresh detail. I keep watching position size and fundamentals.