Gavin Baker - AI Market Jitters - [Invest Like the Best, EP.485]

Episode Summary

Executive Summary: The episode argues that despite a sharp AI stock selloff, the underlying demand for compute is still accelerating across frontier labs, open source, inference clouds, and AI-native companies. The speakers contend that higher GPU prices, long-term agreements, and rising operating cash flow can sustain the buildout, while regulation and credit remain the main risks.

Main Topics: AI demand is accelerating despite market weakness (Priority: 5/5): The hosts repeatedly emphasize that July felt like a major drawdown in AI equities, but on-the-ground demand signals—tokens, GPU pricing, rentals, and lab activity—are all improving rather than slowing. Compute shortages, repricing, and operating cash flow (Priority: 5/5): A central thesis is that installed compute is underpriced versus spot markets, and as contracts roll off, hyperscalers could reprice higher and fund expansion from operating cash flow instead of debt. Open source and inference clouds expand total compute demand (Priority: 5/5): Open source models (e.g., GLM 5.2, Kimi K3) and companies like Fireworks, Modal, and Together are shifting mix away from frontier tokens, but the speakers argue this increases total tokens and compute consumption. Credit, real yields, and capital-cycle risk (Priority: 4/5): The conversation flags rising real yields, widening spreads, and worse-than-expected bond pricing as the main genuine risk if the industry needs debt to finance infrastructure buildout. NVIDIA, LTAs, and the economics of the supply chain (Priority: 4/5): The speakers discuss NVIDIA’s low valuation, its financing and royalty-style business models, and how long-term agreements (LTAs) and supply allocations now determine market share and bargaining power. Policy, PR, and regulatory backlash (Priority: 4/5): The episode argues that AI companies have failed to explain the benefits of data centers and AI to the public, leaving them vulnerable to moratoriums, water/power fears, and broader regulatory resistance. Long-horizon technical breakthroughs and future demand (Priority: 3/5): Continual learning, sample-efficient learning, and SRAM-based accelerators are framed as potentially transformative innovations that could improve ROI and reshape demand patterns over time.

Key Arguments: The selloff in AI names is disconnected from fundamentals; every quantitative signal discussed is accelerating rather than decelerating. Open source is not bad for compute demand because tokens are still tokens: cheaper models may reduce user cost but increase total token volume and total flops consumed. Frontier labs and AI-native companies are still in a compute shortage; no one interviewed had excess GPUs, and many are willing to pay much higher renewal rates. As old contracts reprice to current spot levels, hyperscaler operating cash flow could accelerate enough to fund most or all of the buildout internally. Credit is the key vulnerability: if buildout requires debt and rates/spreads remain elevated, the capital cycle could turn negative quickly. Long-term agreements and supply allocations make breaking contracts strategically dangerous, so buyers are incentivized to honor them even in downturns. NVIDIA’s low valuation looks inconsistent with its financing role, dominance, and embedded optionality from equity stakes, revenue shares, and credit wrappers. Regulation and public perception may be the most underappreciated risks, especially around power, water, data-center siting, and job displacement narratives.

Data Points: AI stock drawdown: 40% to 60% in a month - Describes the July selloff in AI-related equities. Hyperscaler operating cash flow growth: 28 to 32 - Reported acceleration in operating cash flow across Microsoft, Meta, and Amazon. Adjusted operating cash flow including unusual items: 28 to 35 - Includes one-time items and legal/fine expenses. Example GPU rental price: mid-$2 per GPU hour to just under $4 - A company renting a Blackwell cluster expects renewal pricing nearly doubling in seven months. Potential compute price increase example: 100% more for Blackwells - Inference cloud anecdote about contract renewal pricing. Tokens as share of comp spend: 20% to 30%+ - AI-native companies are spending a large share of comp on tokens; one example cited at 30%, with one anecdotal 50% figure. NVIDIA forward P/E: lowest of the last 10 years - As of recording, the market is pricing NVIDIA at a very cheap multiple. AI usage base: 500,000 people - Estimate of current agentic AI users discussed as a tiny share of the global population. Global population reference: 7 to 8 billion people - Used to frame the potential scale-up in AI adoption. SpaceX compute buildout rumor: 8 gigawatts over 18 months - Cited from a public report about SpaceX’s potential data-center/computing expansion. SpaceX revenue at given power assumptions: $50 billion per gigawatt - Used to imply very large revenue potential if the buildout occurs. Consensus estimate for SpaceX next year: $73 billion - Referenced in comparing projected compute monetization to market expectations. Blue-collar job impact: no exact numeric value - Claim that data centers are creating sustained high-paying jobs for electricians, HVAC workers, plumbers, and others. Vendor assessment time reduction: up to 50% - From the Vanta sponsor read, describing automated trust/compliance benefits.

Pivotal Quotes: "The main thing people are saying is the third-party data suggests that the anthropic curve started to go off of its trajectory a little bit." — Speaker: A rare potential negative data point raised during the discussion of accelerating AI demand. "Tokens are not equal, but broadly speaking, all open source taking share does is take margin dollars out of the frontier model layer." — Speaker: Core argument that open source shifts margin, not underlying compute demand. "If the operating cash flow does not continue to accelerate, that would be negative." — Speaker: Defines the key fundamental condition that would invalidate the bullish thesis.

Implications: If the speakers are right, AI infrastructure demand may keep compounding even during equity volatility, favoring compute, memory, and power suppliers. The biggest watch-items are credit conditions and regulation, not demand.