Gavin Baker - AI, Semiconductors, and the Robotic Frontier - [Invest Like the Best, EP.385]

Episode Summary

Executive Summary: Patrick O’Shaughnessy and Gavin Baker discuss how AI is remaking tech, semis, data centers, and investing. Gavin argues scaling laws, infrastructure efficiency, and unique data will determine winners among the Mag 7, while robotics and on-device AI could reshape labor, competition, and society.

Main Topics: AI as an existential race (Priority: 5/5): Gavin says major tech firms now compete in the same AI swim lane and will spend aggressively. Semiconductors and data-center bottlenecks (Priority: 5/5): GPU speed gains outpaced networking, storage, and memory, creating low utilization and high costs. A unified AI efficiency equation (Priority: 5/5): He proposes MAMF, SFU, checkpointing, and PUE as the key framework for model economics. Synthetic data and model scaling (Priority: 4/5): He believes synthetic data works and scaling laws may continue, though no one fully understands why. Inference shifts to the edge (Priority: 4/5): He expects phones to run smaller models locally, with cloud used only when needed. Robotics and autonomy (Priority: 4/5): He sees Tesla FSD and humanoid robots as major near-term disruptions powered by AI. Investing process in an LLM world (Priority: 4/5): He thinks LLMs will favor fundamental investors and move venture toward judgment over pure knowledge.

Key Arguments: Big tech now fights in one AI lane; leaders view losing as existential, not ROI-driven. Scaling laws imply very high marginal costs, but infrastructure efficiency can create huge edge. MFU is the best current lab metric because it captures how well compute is actually used. MAMF, SFU, checkpointing, and PUE explain why some labs will be far cheaper per output. Synthetic data appears to work, so data scarcity may not stop model scaling. Model value may shift from the base model to distribution, unique data, and real-time feedback. Inference will increasingly move to phones because local inference is effectively free. Tesla’s FSD advantage comes from massive real-world video data and scaling compute. LLMs will likely help fundamental investors more than pure quants by improving judgment and programming. Venture and growth investing will hinge more on operational value add and JQ than on simple access.

Data Points: current biggest coherent GPU cluster: 32,000 - Largest coherent cluster mentioned before very recent developments XAI target cluster size: 100,000 GPU cluster - Cluster Elon’s team aimed to make coherent in Memphis GPU training utilization metric: 35 to 40% - Typical MFU reported for many labs Single-GPU software efficiency: 83% - NVIDIA GPUs per the cited MAMF testing AMD Instinct MI300 software efficiency early: 25 or 30% - Initial MAMF when first released AMD Instinct MI300 software efficiency later: 60% - Per the cited stance tests Cloud/cloud-adjacent search behavior: less than 1% - Current global internet market cap founded in two years after Netscape Navigator Meta stock move: down like 80% - Attributed partly to metaverse spend and Apple IDFA changes Tesla autonomy step change: 12.3 - Release where he says almost all human code was removed Tesla next step change: 12.5 - Release that reportedly runs best on AI4 hardware Tesla data source advantage: 100x, 1,000x, 10,000x - Range he uses versus Waymo’s second-largest visual data set Consumer AI IQ uplift example: 100 IQs to 115 - He argues AI can materially raise effective human capability Cloud intelligence pricing examples: $20 a month; $60 a month; $1,000 a month; $10,000 a month - Illustrative tiers for local/cloud intelligence and superintelligence

Pivotal Quotes: "I think it is supremely important for humans that we do not end up in a world where there is just one dominant model." — Gavin Baker: On AI competition and the risk of one model controlling values and information "The GPU without storage, memory, and networking is worthless." — Gavin Baker: On why data-center bottlenecks matter as much as chip speed "what LLMs do, what AI does, is it means the human language is the programming language." — Gavin Baker: On how AI changes software development and investing workflows

Implications: The next winners may be the firms that best convert compute into useful intelligence; listeners should watch data-center architecture, edge inference, and real-world data moats.