GPUs, TPUs, & The Economics of AI Explained | Gavin Baker Interview

Gavin Baker argues reasoning models saved AI during the Blackwell transition gap, and SaaS companies are making an existential mistake by refusing to accept lower AI gross margins.

• Reasoning models (RL with verified rewards + test time compute) bridged an ~18-month gap when pre-training stalled waiting for Blackwell; ARC-AGI went from 8% to 95% in 3 months after OpenAI’s first reasoning model.
• Google trained Gemini 3 on TPU v6/v7 while Nvidia’s Blackwell was delayed, giving Google a temporary pre-training cost advantage; first Blackwell-trained model expected from XAI in early 2026.
• Google is estimated to pay Broadcom ~$15B/year (50-55% margins) for TPU back-end design; Baker calls bringing this in-house economically inevitable as TPU revenue scales toward $30-50B.
• SaaS companies (Salesforce, ServiceNow, HubSpot, Atlassian named) are repeating brick-and-mortar retail’s e-commerce mistake by refusing to accept 35-40% AI gross margins; Baker calls it a life-or-death error.
• When power is the binding constraint, tokens-per-watt dominates TCO; highest-performance chips win regardless of price, giving Nvidia and top-tier silicon sustained pricing power.
• Nuclear can’t be built fast enough in the U.S. due to regulation; Baker sees natural gas and solar as the only realistic near-term power solutions for AI data centers.
• Public nuclear and quantum investment vehicles are not the leading players; Google, IBM, and Honeywell Quantum lead quantum, making current public names poor expressions of the theme.
• Baker’s bear case: in ~3 years, a larger phone may run a pruned frontier model locally, potentially shifting inference off centralized data centers and compressing cloud compute demand.

2025-12-09 · Watch on YouTube