Bearish Into November. Room to Run After: Why Dan Niles Is Watching Hyperscaler Credit Default Swaps

Episode Summary

Executive Summary: Dan Niles argues AI is a real industrial revolution but also a classic bubble: adoption, tokens, and profitability are still improving, yet valuation, leverage, political backlash, and debt-financed capex raise the risk of a sharp unwind. He prefers selective exposure, sees semis and software as increasingly differentiated, and thinks open-source pressure will shift profits from model makers to infrastructure and app layers.

Main Topics: AI as a real revolution, but still a bubble (Priority: 5/5): Niles compares AI to prior industrial revolutions and says the scale of investment is rational for a transformative technology, but that overinvestment is inherent and likely to produce a future drawdown. Semiconductor cycle, valuations, and China risk (Priority: 5/5): He remains constructive on the AI buildout but warns semis are still cyclical and may face a harder reset than many expect, especially if Chinese memory and storage suppliers ramp faster than the market assumes. Token economics and AI profitability (Priority: 5/5): He tracks AI demand through token volume and token pricing, noting token prices have fallen sharply while usage has surged, and that cloud vendors are seeing accelerating revenue and better margins. Political and regulatory risk to AI infrastructure (Priority: 4/5): Niles says data centers are facing rising local and political opposition, which could slow buildout and become a short-term headwind into the midterms. Winner-take-most dynamics in models and software disruption (Priority: 4/5): He thinks open-source and low-cost models will pressure frontier model economics and that value will migrate away from model providers toward infrastructure and software companies with defensible moats. Credit, leverage, and market structure (Priority: 4/5): He emphasizes that debt, leverage, and credit spreads are crucial signals, arguing the AI boom is becoming more fragile as large firms fund capex with more debt and CDS levels rise. Risk management and investing philosophy (Priority: 5/5): Niles stresses downside protection, flexibility, and avoiding leverage. He rejects a pure buy-and-hold mindset and says investors must adapt as facts change.

Key Arguments: AI resembles prior transformative buildouts like railroads and the internet, so overinvestment is normal; the key question is when the bubble breaks, not whether investment is excessive. The current AI bubble is less extreme than the dot-com era because valuations are lower, so any eventual semiconductor drawdown may be severe but probably smaller than 2000-2002. The best near-term indicators are token volume, token pricing, cloud revenue growth, and operating margins; token prices fell about 50% but token usage rose about 2.5x, implying demand is still strong. Public cloud providers are still benefiting: revenue growth accelerated and operating margins improved, suggesting AI usage is translating into better economics. Political backlash against data centers may be a major real-world governor on AI expansion, especially around midterms, even if public opinion improves with education. Open-source and open-weight models are commoditizing lower-end AI use cases, which will squeeze economics for frontier model makers and push value toward infrastructure and application layers. The AI value chain will likely follow winner-take-most dynamics, so not every model company will thrive; OpenAI faces more pressure than Anthropic and Google in his view. Semiconductors remain cyclical, and China’s state-backed push into DRAM/NAND could intensify competition and break the cycle sooner than many investors expect. Debt financing and rising credit-default-swap levels matter because the AI buildout is increasingly capital-intensive and capital markets may not absorb unlimited issuance. Investors should prioritize downside protection because bubbles can persist longer than expected, but when they reverse, losses can be extreme and leverage can be fatal. Software is mixed: security, system-of-record platforms, and some gaming names may be safer, while many point solutions could be disrupted by agentic AI. Software pricing may shift from seat licenses to usage-, token-, and outcome-based models as AI changes how enterprise software is sold.

Data Points: NVIDIA valuation: ~15x earnings - Used as evidence that AI leaders are less expensive than Cisco was during the dot-com bubble. Cisco valuation at dot-com peak: 100x+ PE - Referenced as the classic example of bubble-era valuation excess. Token prices: down ~50% since end of May - Niles said cheaper open-source/open-weight models have compressed what can be charged per token. Token usage: up ~2.5x since end of May - Higher volume has offset falling prices and kept fundamentals healthy. Big cloud revenue growth: 35% to 43% YoY - AWS, Azure, and Google Cloud accelerated from March to June quarter. Big cloud operating margins: up about 2 percentage points - Margins expanded alongside revenue growth, reinforcing the fundamental case. OpenAI and Anthropic annual run rate: $29B to $105B - He used this jump to illustrate how fast AI spending can scale and pressure other software budgets. Software market size: a little over $1T - He contrasted AI spend with the overall software budget pool. Knowledge worker spend: $35T to $50T annually - Used to argue that AI’s long-run spending displacement could be huge even if current model revenues look small. Data center opposition poll: 71% against data centers - He cited a Gallup poll showing local political resistance to new data centers. Poll vs nuclear: 53% against nuclear - Used to show how surprisingly unpopular data centers are relative to nuclear power. Treasury yields: 30-year yield at highest since 2007 - Cited as a sign that higher rates and debt supply are pressuring capital markets. Government debt: $40T - Used to explain why credit conditions are tightening. U.S. GDP: $33T - He compared debt to GDP to emphasize fiscal strain. Deficit level: 6% of GDP - He said this is the highest outside a major war. S&P 500 September seasonality: only month with negative average returns since 1957 - Part of his cautious near-term market outlook. Midterm drawdown statistic: ~10% peak-to-trough from Jul 31 to Nov 9 since 1990 - He used this to support a seasonal/political risk warning. Non-midterm drawdown statistic: ~5% peak-to-trough - Historical comparison to midterm election years. China CXMT DRAM market share: ~8% of global DRAM market - He cited this as evidence China is ramping quickly from a low base. CXMT wafer starts plan: 300,000 to 500,000 by end of next year - Example of aggressive Chinese capacity expansion. Intel DRAM market share historically: 75% - Used to illustrate how dominant incumbents can be displaced in memory semiconductors. NVIDIA revenue guidance: ~70% growth for calendar 2027 - He said the guidance helped confidence but also reminded him of Cisco’s overly optimistic long-term forecasts.

Pivotal Quotes: "There are only two things that drive stocks, right? It's earnings, it's the multiple on those earnings." — Dan Niles: Explaining why valuation, leverage, and fundamentals all matter in the AI bubble debate. "You don't need a Ferrari to go to the corner store to get milk, right? A Ford will work just fine." — Dan Niles: Describing how cheaper open-source models will undercut premium AI pricing in many use cases. "I believe saying, you know, you there's some stocks you just need to buy and hold is completely moronic because you don't know that." — Dan Niles: His core investing philosophy on flexibility and avoiding survivorship bias.

Implications: AI likely has much more room to run, but returns will become more selective as pricing compresses, politics intrudes, and debt/credit tighten. Investors should favor moats, infrastructure, and downside protection over broad beta and leverage.