Gavin Baker - Nvidia v. Google, Scaling Laws, and the Economics of AI - [Invest Like the Best, EP.451]
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this deep dive conversation with Gavin Baker, Managing Partner and CIO of Atreides Management, we explore the intricate dynamics shaping the AI landscape today. (05:28) Baker, who has been covering Nvidia for over two decades, shares his encyclopedic knowledge of the semiconductor ecosystem and the strategic chess match between tech giants. The discussion covers everything from Google's TPUs versus Nvidia's GPUs, the implications of Blackwell chip delays, and the economics of AI token production. (30:13) Baker also presents a fascinating case for data centers in space as the future of AI infrastructure, while examining the broader implications of artificial intelligence on business models and society.
- Core Theme: The strategic and economic battle between major tech companies for AI supremacy, with particular focus on infrastructure, semiconductors, and the evolving competitive landscape.
Speakers
Patrick O'Shaughnessy
Patrick O'Shaughnessy is the CEO of Positive Sum and host of the popular Invest Like the Best podcast. He has built a reputation for conducting in-depth conversations with leading investors, entrepreneurs, and business leaders, focusing on strategies that help people better invest their time and money.
Gavin Baker
Gavin Baker is the Managing Partner and Chief Investment Officer of Atreides Management, with over two decades of experience covering Nvidia and the semiconductor ecosystem. He previously worked as a sector leader for telecom and utilities teams and has developed a reputation as one of the most knowledgeable investors in the technology space. Baker is known for his encyclopedic understanding of AI infrastructure and his passionate, analytical approach to investing.
Key Takeaways
Use Premium AI Tiers for Serious Analysis
Baker emphasizes that many investors are making critical mistakes by evaluating AI capabilities based on free tiers rather than premium offerings. (05:29) He argues that using free AI is "like dealing with a 10-year-old" while premium tiers ($200/month) are like "a fully fledged 30, 35 year old." This distinction is crucial because investment decisions about AI companies are being made based on inferior product experiences. For professionals, this means always testing the highest tier of any AI service before making judgments about its capabilities or market potential.
Scaling Laws for Pre-Training Remain Intact
Gemini 3's release provided crucial validation that scaling laws for pre-training continue to hold, which Baker calls "very important" since no one understands how or why these laws work. (08:27) He compares our understanding to ancient civilizations' precise measurements of the sun without understanding orbital mechanics. This empirical observation drives massive infrastructure investments, and its continuation means that larger, more expensive models will continue to be meaningfully better. The implications are profound for capital allocation decisions across the tech industry.
AI Can Automate Anything That Can Be Verified
Drawing from Andre Karpathy's insight, Baker explains that while software can automate anything you can specify, AI can automate anything you can verify. (10:32) This means functions with clear right/wrong outcomes—like accounting (books must balance), sales (conversion or no conversion), and customer support (escalation or resolution)—are prime targets for AI automation. This framework helps identify which business functions will be transformed first and provides a roadmap for both investors and business leaders to understand where AI will create the most immediate value.
Power Constraints Favor Advanced Technologies
When watts become the primary constraint rather than cost, Baker argues this creates massive advantages for the most advanced computing technologies. (60:44) If a $50 billion advanced data center generates $25 billion in revenue while a $35 billion ASIC-based center only generates $8 billion, the economics clearly favor the advanced option regardless of upfront costs. This dynamic means that as long as power remains a bottleneck, the best technologies will win irrespective of price, creating significant pricing power for cutting-edge solutions.
SaaS Companies Must Embrace Lower Margins or Risk Obsolescence
Baker warns that SaaS companies are repeating the same mistake brick-and-mortar retailers made with e-commerce by refusing to accept AI's lower gross margin structure. (68:39) While traditional SaaS enjoys 70-90% gross margins, successful AI companies operate at around 40% margins but achieve profitability faster due to fewer human employees. Companies trying to preserve 80% gross margins are "guaranteed" to fail in AI, while those willing to run agent strategies at sub-35% margins can leverage their existing customer bases and cash flow advantages over AI-native startups.
Statistics & Facts
- Google has been the lowest cost producer of tokens, and Baker estimates that if TPU spending reaches $30 billion by 2027, Google would pay Broadcom approximately $15 billion in gross margins (50-55%) for backend semiconductor design and manufacturing. (19:47)
- In space, solar energy provides six times more irradiance than on Earth due to being 30% more intense and available 24 hours per day, making it the lowest cost energy source available in our solar system. (55:30)
- According to Baker, XAI processes 1.35 trillion tokens compared to Google's 800-900 billion and Anthropic's 700 billion on OpenRouter, indicating XAI's strong market position despite being a newer player. (49:23)
