Ep20. AI Scaling Laws, DOGE, FSD 13, Trump Markets | BG2 w/ Bill Gurley & Brad Gerstner

Episode Summary

Executive Summary: The discussion centered on whether AI pre-training scaling is slowing, and what that means for NVIDIA, model labs, and the broader AI stack. The speakers argued that while pre-training gains may be decelerating, AI progress is shifting to post-training, inference-time reasoning, memory, actions, and better product design. They also covered Tesla FSD, robo-taxi prospects, regulatory reform under DOGE, and a constructive but selective 2025 market outlook.

Main Topics: AI pre-training scaling debate (Priority: 5/5): The speakers examined reports that OpenAI, Google, Anthropic, and others are missing pre-training targets, suggesting that LLM pre-training may be nearing a plateau or slowing in its second derivative. Shift from pre-training to other scaling vectors (Priority: 5/5): They argued that even if pre-training slows, progress can continue through post-training, inference-time reasoning, data quality, multimodality, and larger context windows. Implications for NVIDIA and infrastructure demand (Priority: 5/5): A major thread was whether a pre-training slowdown would hurt NVIDIA demand. The conclusion was that it likely would not, because inference, larger contexts, and new product use cases can still drive compute growth. Consumer and enterprise AI product expansion (Priority: 4/5): They emphasized memory, voice, and actions as the next major product unlocks, increasing utility and token usage across consumer and enterprise workflows. Tesla FSD, robo-taxi, and autonomy regulation (Priority: 5/5): The conversation shifted to Tesla's FSD 13, potential safety gains, robo-taxi deployment, and a proposed national regulatory framework that could accelerate deployment while raising liability questions. DOGE, government efficiency, and fiscal reform (Priority: 4/5): They discussed the new Department of Government Efficiency, framing it as a potentially historic effort to cut waste, reduce regulation, and balance the federal budget. 2025 market outlook and dispersion (Priority: 4/5): The speakers ended with a constructive view on 2025: lower taxes, deregulation, and AI tailwinds support tech, but tariffs, consumer stress, and high valuations argue for a stock-picker's market rather than a broad index melt-up.

Key Arguments: Pre-training may be slowing, but that does not imply AI is hitting a wall; other scaling vectors can offset it. The market previously assumed linear or exponential model improvement from ever-larger pre-training runs, so any slowdown is a meaningful shift. NVIDIA remains well-positioned because inference, post-training, and larger context windows can also require large compute clusters. Memory, voice, and actions can materially increase AI utility even without major core model breakthroughs. Robo-taxi economics may be unlocked by markets with dense Tesla ownership and favorable safety data, rather than everywhere at once. A national autonomy framework could help the industry, but litigation and liability limits may also be necessary. DOGE could reduce waste and regulation enough to improve growth and help balance the budget without draconian cuts. 2025 is likely to feature major dispersion: winners in AI/automation and losers among exposed consumer, retail, and legacy OEM businesses.

Data Points: Government agencies in Friedman clip: 14 - Referenced while discussing the philosophy behind abolishing or retaining federal departments. Azure run rate: $66 billion - Mentioned as Microsoft's Azure revenue base, with expected acceleration. Azure growth rate: 34% - Used to illustrate strong software and cloud growth despite the AI debate. Federal revenues: About $5.2 trillion - Used in a back-of-the-envelope federal budget analysis. Federal spending: About $7 trillion - Used to illustrate the current deficit and why cost cuts matter. Federal deficit: About $2 trillion - Derived from spending minus revenue in the budget discussion. Federal spending in 2019: $4.5 trillion - Referenced to show how much spending has expanded since COVID. Target stock reaction: Down 25% - After the company missed numbers and took down guidance. Snowflake after-hours reaction: Up 20% - Cited as an example of AI beneficiaries with low expectations and margin upside. Snowflake customer usage: 3,200 customers - Out of roughly 10,000 enterprise customers using AI products daily. Tesla cars on the road: About 7 million - Used to estimate the base of vehicles available for robo-taxi participation. Tesla Hardware 4 vehicles: About 2.5 million - Estimated subset required for running the newer autonomy stack. FSD improvement at start of year: 100x to 1,000x range - Described as improvement in miles per critical disengagement across FSD versions. FSD 13 projected MPCI: Around 25,000 miles per critical disengagement - Estimated launch performance for Tesla's next FSD version. Waymo MPCI estimate: Around 20,000 miles - Used as a comparison to Tesla's projected safety performance. Human accident frequency: 1 every 500,000 miles - Used to frame the threshold for robo-taxi viability. Potential 2025/next-year robo-taxi milestone: Q2 - Suggested window for launching robo-taxi in a couple of markets. S&P valuation: 24x earnings - Used to argue broad markets are not cheap heading into 2025. NASDAQ valuation: 29x earnings - Used alongside S&P valuation to support a selective rather than aggressive market stance. Microsoft income margin expansion: 1,000 basis points - Cited as evidence that AI and software can drive substantial profitability expansion over time.

Pivotal Quotes: "scaling laws are not actually laws of the universe. But they are simply empirical regularities." — Dario Amodei (quoted by speaker): Used to frame the debate over whether AI pre-training improvements are slowing or just changing form. "I definitely think the second derivative is slowing, right? But you're still seeing gains." — Speaker: Summarized the view that pre-training progress is decelerating without stopping outright. "This could be a once-in-a-generation reset in the balance of power between the individual and the state." — Speaker: Describing the potential impact of DOGE and government reform.

Implications: AI investment is likely to rotate from pure pre-training scale to inference, memory, actions, and product quality. Expect big winners and losers across chips, cloud, autonomy, and legacy industries, plus a more selective 2025 market.