The AI Model That Tanked the Stock Market

Episode Summary

Executive Summary: The episode centers on the market sell-off triggered by DeepSeek, a Chinese open-source AI model that appears to match leading U.S. reasoning models at far lower apparent training cost. Guest Zvi Moshevitz argues the headline $5.5 million figure is real but incomplete, and that the bigger takeaway is not AI demand destruction but cheaper, broader AI usage that likely boosts compute demand over time.

Main Topics: DeepSeek and the AI-stock sell-off (Priority: 5/5): The hosts frame Monday’s market rout as a reaction to DeepSeek, a high-performing Chinese model that sparked fears about the durability of U.S. AI leaders and chipmakers, especially NVIDIA. The true cost of training DeepSeek (Priority: 5/5): Zvi explains that the reported $5.5 million refers to training V3, not the full end-to-end effort, and that DeepSeek still required major spending on data, engineers, infrastructure, and optimization. Open-source AI and strategic risk (Priority: 4/5): The conversation examines why DeepSeek’s open-source approach matters, including concerns that increasingly capable models could be freely available and harder to control. Jevons paradox and compute demand (Priority: 5/5): The discussion argues that efficiency gains may increase overall AI usage rather than reduce demand, because cheaper inference encourages more queries, more applications, and more total GPU consumption. Competitive impact on OpenAI, Anthropic, Meta, and Google (Priority: 4/5): The guests assess how DeepSeek pressures U.S. AI players: OpenAI faces direct reasoning-model competition, Anthropic may be constrained by compute, Meta’s open-model advantage is weakened, and Google remains underappreciated. Moats, scaling laws, and the 'bitter lesson' (Priority: 4/5): The episode argues that AI moats are weaker than in search because everyone trains on similar internet-scale data and can copy techniques quickly, making model lead times harder to defend.

Key Arguments: DeepSeek’s reported $5.5 million training cost is real but only part of the total investment; the broader program still required hundreds of millions in infrastructure, engineering, and optimization. The model matters because DeepSeek shared enough technical detail that competitors can copy many of its efficiency tricks, lowering the cost of follow-on model training. Open-source release of increasingly capable models raises strategic and existential concerns because powerful AI becomes accessible to anyone. Efficiency gains are likely to increase, not decrease, total GPU demand because lower costs expand the set of feasible AI use cases and increase inference-time usage. DeepSeek weakens Meta’s position in open models because its own open-source model appears better than Llama-based alternatives. AI model moats are thinner than search moats because the underlying data, techniques, and outputs are easier to replicate and iterate on. Google’s AI efforts are arguably underrecognized by the market compared with OpenAI and others, despite competitive advancements in Gemini.

Data Points: NVIDIA intraday stock move: down 17% - Cited as a major market reaction to DeepSeek-related fears NVIDIA market value wiped out: $560 billion - Referenced jokingly as the scale of the sell-off tied to DeepSeek DeepSeek V3 training cost: $5.5 million - The technical paper’s reported cost for training V3, not the full R1 system DeepSeek training data scale: 15 trillion tokens - Used to explain the efficiency and scale of DeepSeek’s training process Meta announced AI spend: $65 billion - Mentioned as a recent capex commitment, viewed more positively by markets Meta stock move: +1.1% - Observed intraday while discussing how markets reacted to AI spending OpenAI reasoning model: O3 can think for many minutes - Used to illustrate increasing inference-time compute requirements OpenAI query cost potential: hundreds or even thousands of dollars per query - Hypothetical upper-end inference costs for highly deliberative models Chip access: H800s - DeepSeek reportedly trained under export-control constraints using these chips

Pivotal Quotes: "it’s bad in markets when all the headlines are about standard deviations" — Joe Wisenthal: Opening remark about the severity and abnormality of the sell-off "we’re in deep seek, yeah" — Joe Wisenthal: Wordplay on the DeepSeek-driven market panic "I’m definitely a Jevons paradox bro right now" — Zvi Moshevitz: Arguing that efficiency gains in AI will increase, not reduce, overall compute demand

Implications: DeepSeek suggests AI competition is accelerating, model gaps may be easier to close, and cheaper inference could expand total AI usage. Investors should expect more volatility in chip and AI stocks, not less.