Why AI Stocks Crashed, OCP Conference Takeaways, And What Were Buying
At the beginning of the week, Michael and Kawz had just arrived at the Open Compute Project summit to get the latest information on AI infrastructure. We were within a fraction of all time highs on the Nasdaq and S&P 500. The SOX sat about 14% off its own peak but had climbed roughly 25% from the July lows and nothing about the setup suggested we’d revisit those levels anytime soon. Then a series of events compressed into five trading days. By Thursday, the SOX had posted seven straight down sessions and Korea’s KOSPI tripped a circuit breaker after falling 7.6%. Some of the neoclouds and optical names dropped close to 25% on the week. Hike odds for October jumped to around 53% after the inflation prints, then faded to 87% probability of maintaining rates by Today. Then every company that reported this week guided up. ASML narrowed its full year range higher and TSMC beat on revenue, and guided Q4 above consensus. Anthropic grew 45% in a single quarter and posted its second straight operating profit. The supply chain said demand is accelerating but the stock market said it doesn’t care. That disconnect is the whole week. We think the selloff was a mix of both, the macro outlook changed as inflation moved the rate path, but at the stock level it was mostly sentiment. However, CoreWeave is the one exception. Before we begin, this Substack is for informational and entertainment purposes only and does not constitute financial advice. Past performance is not indicative of future results. Never invest more than you can afford to lose. Our writers may hold positions in the securities discussed and may buy or sell them at any time without notice. Table of Contents: OpenAI’s $20B Revenue Gap Hock Tan’s Casino Quote September CPI September PPI Bankruptcy Fears Semi Earnings and GPU Demand Anthropic’s Q3 Numbers Warsh in Bangkok What to Watch [1] OpenAI’s $20B Revenue Gap This story broke a week before our coverage window, but it set the tone for everything that followed. On Sep 29, Axios reported that OpenAI was on track to generate annualized revenue of nearly $70B. Nine days later, on Oct 8, the Financial Times reported that OpenAI had told investors its ARR was approaching $50B at the end of September. The Nasdaq dropped 1.25% that same trading day. The $20B difference came down to how revenue was measured, that $68 to $70B figure counted sales made through cloud partners, the way Anthropic reports its revenue, and the FT’s $50B is OpenAI’s own count, which leaves those sales out. OpenAI did $13B in FY25 revenue and the run rate now, a year later, sits at $50B. Whether that number validates the growth story or falls short depends entirely on which figure investors had been underwriting, but anyone benchmarked against $70B had to mark down by nearly 30% overnight. [2] Hock Tan’s Casino Quote Monday was the first session where the selling broadened out, the SOX dropped 2.7%, optical, and quantum names took the worst of it: Ciena fell 10.5%, Lumentum 9.8%, IonQ 9.0%. On Monday, Broadcom CEO Hock Tan spoke at the DBS Asia Future Forward conference in Singapore. He was asked about cheaper AI models and whether falling costs would shrink chip demand. He called OpenAI’s $50B run rate “by itself, not that bad” and said capital, rather than chips, was the biggest constraint facing the buildout. Then came the line that moved the tape: the market had “set very extreme expectations only to find out that it’s almost like they are in a casino.” The whole finX and the market read that as Broadcom’s CEO calling AI investors gamblers. However, Tan was criticizing the way expectations get built around AI revenue targets, where the market anchors on an aggressive number and then panics when reality comes in below it. His actual view on chip demand was that capital availability is the bottleneck. That framing is constructive for semis, but the headline stripped away the nuance and the tape reacted to the word “casino.” [3] September CPI Tuesday was ugly in Seoul before any US data had printed, because the KOSPI fell 5.46%. After that the inflation numbers arrived, the BLS (Bureau of Labor Statistics) released September CPI on Wednesday Oct 14. → Headline CPI came in at 3.7% YoY against a 3.6% consensus. → Core CPI printed +0.3% MoM, above the 0.2% expected. Small misses on both readings, but enough to keep the Fed conversation alive. Wednesday was the worst session of the week. The SOX fell 4.30% and CoreWeave dropped 15.5% due to bankruptcy fears. Dell lost 11.3%, with a Barclays downgrade on Dell, HPE, and NetApp landing the same morning. Neoclouds gave up around 10% across the whole board and the S&P 500 fell 1.11%. [4] September PPI PPI landed Thursday morning and was even harder to dismiss than CPI. → Headline PPI rose 6.3% YoY versus a 5.3% forecast. → MoM PPI came in at +1.3%, more than double the 0.5% expected. The BLS called the monthly increase the largest since March 2022 and the yearly figure the biggest since December 2022. Energy drove the overshoot at +10.8% for the month. The connection between hot inflation prints and AI infrastructure stocks is simple because higher rates mean higher borrowing costs. Companies financing tens of billions in GPU clusters and data centers get repriced when the yield curve moves against them. The 10-year Treasury stayed above 5% through the week and for a company like CoreWeave, which carries roughly $35B in debt, every basis point matters on the refinancing schedule. [5] Bankruptcy Fears CoreWeave’s CDS spread hit a record 1,194 bps on Thursday, meaning insuring its debt now costs about 12% of the amount covered, per year. The number floating around FinX was a 15% yield on CoreWeave’s bonds and Kawz broke this down well on his feed. The 15% is the market yield on the 2032 bond, which was locked in at a 9.625% coupon. CoreWeave keeps paying 9.625% regardless of where the bond trades in secondary markets. A new buyer earns 15% because the bond price has fallen. Across all its debt, the blended cost sits around 8.4% and only about 4% of the company’s roughly $35B in debt, a $1.3B term loan, actually costs CoreWeave around 15%, and it matures March 2028. Most of the borrowing was locked in at lower rates before the recent move in Treasuries. The refinancing risk is there but it’s concentrated in that small slice maturing in 2028, and the company has time before that becomes a liquidity event. We can’t confirm whether forced selling happened. But the names in VARA’s portfolio overlap with the stocks that sold off hardest on Wednesday. If VARA or similarly positioned funds had to reduce exposure, that selling pressure would’ve landed on an already fragile tape. We saw a similar playbook in July when SALP’s forced liquidation drove the selloff that took the SOX to those lows we’d just climbed 25% back from. VARA AI’s fund exposure vs drawdown [6] Semi Earnings and GPU Demand Two of the most important companies in the semiconductor supply chain reported this week, right in the middle of the selloff and both confirmed that the demand picture hasn’t changed. ASML reported Q3 on Wednesday. Revenue came in roughly in line with consensus, but gross margin came in at 57.5%, above guidance, driven by high margin productivity upgrades across the EUV and DUV installed base. The company narrowed its FY guide upward to €44 to €45B and pointed Q4 toward €14.2 to €15.2B, a 22 to 30% sequential jump. Around €500M of net income came from a gain on ASML’s Mistral stake, so strip that out when comparing to prior quarters. The earnings commentary was even more important for our long term horizon than the headline numbers. The CEO Christophe Fouquet said: “supply shortages across advanced logic and memory nodes continue to increase in size and in time horizon.” Customers are pulling in orders for 2028 and asking to place orders for 2029 and 2030. CFO Roger Dassen on why they haven’t accepted those yet: “We do not want now to accept... orders from customers A, B, and C... and then to tell customer D who has not... given orders yet for 2029 and 2030... Sorry. We are sold out.” ASML is now looking into expanding 2028 capacity beyond the 30% growth it flagged last quarter. Fouquet said cycle time reductions and footprint optimization are giving them room, and he said: “we do not see Zeiss (the company that makes the optical systems inside ASML’s EUV machines) today as a bottleneck for the capacity expansion we are discussing, even beyond the 30%.” Then, TSMC reported Q3 on Thursday. Gross margin was 66.4%, with the sequential dip driven by 2nm ramp dilution. The company guided Q4 to $52.3 to $54.3B, above consensus, and reiterated a 2026 capex plan of $60 to $64B. Chairman C.C. Wei on demand: “the demand is very extremely robust, okay? And we did not see any sign of slowdown.” TSMC had previously guided its AI accelerator revenue CAGR at mid to high 50s percent from 2024 to 2029 and when asked for an update, Wei said “it probably will be higher.” Furthermore, TSMC is expanding at an unprecedented pace with nearly 20 fabs worldwide in progress. Wei closed with: “we have no intention to leave food on the table for others, and we want to capture every piece of business possible.” GPU rental pricing backs up the demand picture that both ASML and TSMC presented, SemiAnalysis data shows B200 1-year contract rates spiking toward $6/hr in October, up from around $4.50 earlier in the year. H100s hold steady around $3/hr. A100s sit near $1.75, up from around $1.35 in February. [7] Anthropic’s Q3 Numbers Anthropic’s Q3 numbers leaked Wednesday via the Financial Times and Reuters. Revenue hit $16.8B for the quarter, up 45% from Q2, with an 11% operating margin, which was the company’s second straight quarterly profit. An analyst who attended a private Anthropic event puts core annualized revenue at about $71B exiting September, excluding distillation and platform partner revenue that’s rolling off. If monthly growth holds near the 13% pace it ran through Q3, core ARR crosses $100B by December. That puts Anthropic’s September monthly revenue at around $6.2B, ahead of OpenAI’s roughly $4.2B. CNBC floated a $2T valuation for Anthropic’s pre-Thanksgiving IPO and now other sources put it closer to $1.5T. [8] Warsh in Bangkok On Friday, Fed Chair Kevin Warsh spoke at the IMF meetings in Bangkok. He offered no signal on October: “I don’t prejudge meetings.” He described a global environment of “synchronous tightening” across central banks, meaning the Fed isn’t the only one keeping rates elevated. On AI, he called it a moment of “optimism, but optimism leavened by the seriousness of purpose.” We still have no idea what that means. Goldman already expects a pause in October with the next hike in December, a call it made at the start of the month. For anyone positioned in rate-sensitive AI infrastructure names (most of them), that December hike probability is the number to watch from here. [9] What to Watch We think that during this week the market repriced the discount rate. Hot inflation was a fundamental reason to pay less for AI growth, but the earnings showed that growth is intact. CoreWeave is the exception, where the credit market is pricing actual distress for a business that has to keep borrowing to grow. These are the most important dates to look at closer in the upcoming weeks: Oct 17 - Fed blackout begins, no Fedspeak until after the decision. Important for basically everyone in the market since rate expectations are frozen until Oct 28. Oct 21 - Meta, Texas Instruments, Lam Research earnings. Important for investors in companies spending at scale and for measuring chip demand across analog and fab equipment. Oct 28 - FOMC rate decision and Microsoft earnings. The rate call is the most important event of the month. Microsoft’s results will give us a read on cloud AI ROIC and enterprise adoption trends. Oct 29 - Intel earnings and PCE inflation print. Intel gives us the first major look at CPU demand and how the agentic AI trade is translating into silicon orders. PCE is the Fed’s preferred inflation measure and will shape December hike expectations. Nov 3 to 5 - AMD, Arm, Coherent, Lumentum earnings. The biggest names in CPUs, semiconductor IP, and optical networking all reporting in the same window. These days cover the full AI supply chain from compute to connectivity. [10] What We’re Adding + Stocks To Watch + OCP Recap This past week Michael and Kawz had the chance to attend the Open Compute Project Summit. If you’ve never heard of it, OCP is basically where the people who build AI data centers go to agree on how to build them. Meta started the Open Compute Project back in 2011 to open source its server and data center designs, and today it’s the group that sets the standards Microsoft, Google, Meta, Oracle and the rest of the hyperscalers actually use. This year’s summit ran October 12 to 15 at the San Jose Convention Center under the theme “Scaling Innovation for the AI Era,” and it sold out with no onsite registration (it’s moving to Moscone in San Francisco next year because it’s outgrown San Jose). For us this is the closest thing to seeing the AI buildout’s shopping list in person, since the bottlenecks hyperscalers are complaining about here tend to show up in supplier earnings a few quarters later. There were a few major insights we took away from the event and a few trades we’re taking afterwards. Read more
View original →