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

What It Covers

Gavin Baker analyzes the AI infrastructure race between NVIDIA and Google, explaining how Blackwell chip delays, TPU advantages, scaling laws for pretraining, and the economics of token production shape competitive dynamics among frontier AI labs.

Key Questions Answered

• •Blackwell Product Transition: NVIDIA's Blackwell chips require liquid cooling, 130 kilowatts per rack versus 30 for Hopper, and weigh 3,000 pounds versus 1,000 pounds. This eighteen-month delay would have stalled AI progress without reasoning models bridging the gap until deployment scaled in late 2024.
• •Google's Cost Advantage: Google operates as the lowest cost token producer using TPU v6 and v7 chips, running AI at negative 30% margins to extract economic oxygen from competitors. This advantage disappears when Blackwell-trained models deploy, fundamentally changing strategic calculations across the industry.
• •Reasoning Model Economics: Reinforcement learning with verified rewards and test time compute create multiplicative scaling laws beyond pretraining. These enable the flywheel effect where user interactions generate data to improve models, similar to how Netflix and Amazon achieved increasing returns to scale over decades.
• •SaaS Margin Trap: Application software companies preserve 80% gross margins while AI natives run agents at 40% margins, repeating brick-and-mortar retailers' ecommerce mistake. Companies like Salesforce and ServiceNow must accept lower margins or face displacement by venture-funded competitors accessing their data through APIs.
• •Data Centers in Space: Solar energy in space provides six times more irradiance than Earth with no battery costs, free cooling via radiators, and faster laser communication through vacuum than fiber optics. Starship launch economics make this viable for inference workloads within three to four years.