Invest Like the Best with Patrick O'Shaughnessy · Colossus | Investing & Business Podcasts

Gavin Baker - Nvidia v. Google, Scaling Laws, and the Economics of AI

·1 hr 26 min·5 clips

Gavin Baker reveals how reasoning models bridged an 18-month AI progress gap when Blackwell chips were delayed.

Patrick O'Shaughnessy interviews Gavin Baker, a technology investor known for encyclopedic knowledge of the AI ecosystem, in what Baker describes as a process of keeping up with an extremely fast-moving field: using the highest-tier paid models rather than free versions, following the roughly 500-1000 people on the planet who are genuinely at the cutting edge of AI research, and using AI itself to process and discuss podcast content. Baker opens with a discussion of Gemini 3's significance: its main contribution was confirming that pre-training scaling laws remain intact, an empirical observation that Baker analogizes to ancient Egyptians precisely measuring the sun's movement without understanding orbital mechanics. He explains that after NVIDIA's Hopper generation reached its coherence limit at around 200,000 GPUs, there should have been no AI progress until Blackwell arrived -- but Blackwell was delayed due to extraordinary product complexity, requiring data centers to shift from air-cooled to liquid-cooled racks, tripling rack weight, and quintupling power draw. Reasoning models -- reinforcement learning with verified rewards and test-time compute -- filled that 18-month gap and are now two new scaling laws stacking on top of the confirmed pre-training scaling laws. Baker argues Google's temporary position as the lowest-cost token producer has been a rational strategy to drain economic oxygen from the AI ecosystem, but this calculus changes once Blackwell models begin training and then shift to inference, making NVIDIA customers the new low-cost producers. He describes the Broadcom situation as economically unstable: Google is currently paying Broadcom roughly $15 billion for the back end and management of TPU production, and as TPU revenue grows toward $30-50 billion annually, the incentive to bring that work in-house grows proportionally. Baker identifies what he sees as the first signs of AI ROI in public company financials: C.H. Robinson, the freight forwarding company, printed a 20% earnings surprise driven by AI-enabled quote generation jumping from 60% of inbound requests to 100%, with response time dropping from 15-45 minutes to seconds. He also flags the trend in venture-backed companies showing higher revenue per employee than at the same revenue level two years ago. Baker argues the most important near-term AI product development is not more intelligence but more usefulness: context windows large enough to hold all of a user's preferences, past interactions, and institutional knowledge, enabling models to function as true personal assistants that can plan a vacation with awareness of each family member's preferences. His bear case for AI infrastructure demand is edge AI: if a pruned model running at 30-60 tokens per second on a next-generation smartphone proves good enough for most everyday uses, the incentive to query expensive cloud-based frontier models diminishes. The conversation also covers robotic startups, the advantage of verified-reward reinforcement learning for tasks with right-or-wrong outcomes, and the investing landscape of the 490 S&P 500 companies outside the top 10 mega-caps.

As heard by us

A measured look at AI as a business-system shift, from cloud comparisons to concrete signs of enterprise productivity gains.

The episode treats AI less like a slogan and more like an economic systems story, keeping attention on the machinery underneath. It is at its best when it moves from broad optimism to concrete signs of change: venture-backed companies hiring fewer people for a given revenue…

Read the full review in PlayNext →

Why you'd press play

You want the AI boom explained as economics, not noise.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Invest Like the Best with Patrick O'Shaughnessy