Gavin Baker - Nvidia v. Google, Scaling Laws, and the Economics of AI - [Invest Like the Best, EP.451]

Episode Summary

Executive Summary: Patrick O'Shaughnessy and Gavin Baker dissect the AI arms race: model progress, chip and data-center economics, and how these shifts reshape competition among Google, NVIDIA, OpenAI, Anthropic, Meta, xAI, and startups. Baker argues AI progress is being driven by post-training, test-time compute, and infrastructure advantages that are now visible in public-company financials.

Main Topics: Frontier-model progress and scaling laws (Priority: 5/5): Baker says pre-training scaling still holds, while post-training and test-time compute drove recent leaps. NVIDIA, Google TPU, and the chip race (Priority: 5/5): He frames AI as a competition between GPUs and TPUs, with Blackwell and Rubin shifting the balance. Economics of AI and token production (Priority: 5/5): Low-cost token production is now strategic, because pricing, margin, and capital access decide winners. Reasoning and verifiable rewards (Priority: 4/5): He argues AI improves fastest where outcomes are checkable, enabling reinforcement learning loops. Enterprise adoption and ROI (Priority: 4/5): Baker says ROI is already visible in public-company results and will spread as AI automates core functions. Future bottlenecks and new infrastructure (Priority: 4/5): He highlights power, DRAM, space data centers, and rare earths as the next constraints and breakthroughs. Investing philosophy and origin story (Priority: 3/5): The closing turns personal, tracing his investing mindset to history, current events, and competition.

Key Arguments: Pre-training scaling laws still hold; Gemini 3 confirmed that empirical pattern. Recent AI gains came from post-training RL with verified rewards and test-time compute. Google's low token costs give it a temporary strategic edge by squeezing rivals' economics. Blackwell delays mattered because reasoning bridged the gap until next-gen chips arrived. The first Blackwell-trained model is likely in early 2026, with xAI positioned first. AI value is strongest where outcomes are verifiable, like sales, support, accounting, and models. Enterprise ROI is already positive; major GPU spenders report higher ROIC than before. SaaS firms will have to accept lower gross margins to compete with AI-native agents.

Data Points: Expense reviews automated by Ramp AI: 85% - Ad read claims Ramp automates most expense review work Expense review accuracy: 99% - Ad read claims Ramp’s AI review accuracy Company savings from Ramp: 5% - Ad read claims Ramp saves companies this amount AI tokens processed by OpenRouter (xAI): 1.35 trillion tokens - Baker cites token volume as a rough indicator of model usage AI tokens processed by Google: 8 or 900 billion - Baker compares Google’s recent token volume to competitors AI tokens processed by Anthropic: 700 billion - Baker cites recent usage scale in the model race Customer support done by AI at some tech-forward companies: 50% plus - He says AI already handles over half of support in some companies C.H. Robinson earnings move: 20% - He attributes the stock reaction to AI-driven productivity gains CH Robinson inbound quote coverage before AI: 60% - He says they used to quote only part of inbound requests SaaS gross margins AI-native companies may run: sub 35% - He argues incumbents must accept much lower margins to compete Potential AI assistant speed on phones: 30, 60 tokens per second - Used as the edge-AI bear-case benchmark Model evaluation example: 0 to 8% to 95% in three months - He references ARC AGI progress after reasoning models Rack power jump from Hopper to Blackwell: 30 kilowatts to 130 kilowatts - Illustrates the infrastructure shock of new GPU generations Rack weight jump from Hopper to Blackwell: 1,000 pounds to 3,000 pounds - Illustrates cooling and structural requirements

Pivotal Quotes: "Anything you can verify, you can automate." — Gavin Baker: Explaining why reasoning and RL thrive in domains with clear right-or-wrong outcomes "I think we need to shift from getting more intelligent to more useful." — Gavin Baker: Discussing the next phase of AI adoption beyond benchmark progress "Whatever AI needs to keep growing and advancing, it gets." — Patrick O'Shaughnessy: Closing reflection on how constraints keep getting solved just in time

Implications: The next phase will be judged less by benchmark wins than by real workflow capture, so investors should watch margins, deployment speed, and enterprise adoption closely.