#559. All-in: SpaceX, Recursive AI Self-Improvement, Nvidia’s Massive Profits, and Why America Is Starting to Fear AI

📝 Episode Summary

This episode clones: Silicon Valley’s top-tier tech and macro commentary podcast *All-In Podcast* — [SpaceX’s $2T Case, Nvidia’s Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?](https://www.youtube.com/watch?v=HGbA6ze0_3M)

【This episode contains edits and abridgments】

This episode delivers a high-density discussion spanning AI, space, semiconductors, macroeconomics, and geopolitics. Jason Calacanis, Chamath Palihapitiya, David Friedberg, and guest Gavin Baker begin with Andrej Karpathy’s move to Anthropic, debating whether recursive self-improvement could usher AI models into a “new Moore’s Law” era. They also deeply examine why American society is growing fearful of AI—including job displacement, regulatory moats, poor communication by tech CEOs, and AI’s real-world value in public safety, autonomous driving, healthcare, and manufacturing.

Midway through, the panel dissects SpaceX’s astonishing IPO valuation potential: Starlink has become a cash machine, xAI is rapidly catching up to frontier models, the so-called “Elon Web Services” could emerge as a new cloud computing giant, and space-based data centers and orbital compute are viewed as civilization-level infrastructure backups. They then analyze Nvidia’s blowout earnings report, the GPU vs. ASIC competition, valuation misalignments among AI chip companies, and why GPUs’ financiability and long-term lifespans may rescue neocloud providers. Finally, the conversation turns to macro markets and geopolitics: high debt, high interest rates, the Strait of Hormuz, semiconductor export controls, and the Taiwan issue—collectively forming a globally complex chessboard that is both perilous and full of opportunity.

👨‍⚕️ Guest

Gavin Baker, Founder and CIO of Atreides Management, focuses long-term on investments in technology, internet, semiconductors, AI, and high-growth companies. A frequent guest on *All-In Podcast*, he is renowned for his deep insights into AI infrastructure, semiconductor cycles, SpaceX, Nvidia, and global markets.

⏱️ Timestamps

00:00 Opening & Episode Overview

The Next Acceleration in AI

01:32 Karpathy Joins Anthropic: A Legend Returns to the Frontlines

03:00 Recursive Self-Improvement: What Does It Mean for Claude to Improve Claude?

04:11 Chamath: Karpathy and the Google Fellows–Style Genius Culture

05:53 Gavin: Anthropic’s Profitability, LLM ARR, and Early AI ROI Signals

08:04 When Will Super-Recursion Arrive? Will AI Improving AI Surpass Human Engineers?

08:56 Friedberg: Small Model Networks, Architectural Redesign, and Falling Cost per Token

09:57 Gemini Nano Enters Chrome: Local Small Models, Privacy, and User Value Trade-offs

The Narrative War Around AI

10:40 Chamath: Don’t Demonize AI—Tell Stories Where End Users Truly Benefit

13:10 Gavin: The Tech Industry Has a Responsibility to Advocate for AI’s Positive Potential

14:00 Regulatory Moats: Why Model Company CEOs Amplify AI Risk Narratives

16:15 A Rare-Disease Father’s Story: How LLMs Transformed His Child’s Life

18:04 Friedberg’s Extended Answer: Why Young People Are Starting to Boo AI

18:53 Technology, Power Imbalance, and Anti-Humanism: The Deep Psychology Behind AI Backlash

21:10 External Actors, Technology Diffusion, and the AI Arms Race

24:04 Can We Slow Down AI? Autonomous Driving, Robot Taxes, and Retraining

AI Deployment, Regulation, and Societal Conflict

25:15 Chamath: Before Protecting Jobs, Ask Whether Workers Actually Want Them

26:43 AI Executive Order Withdrawn: Should Frontier Models Be Tested Before Release?

28:11 Gavin: Government Power Is a One-Way Ratchet; Self-Regulation and Judicial Accountability Still Exist

30:38 Autonomous Driving Safety: Will Cities Banning Waymo Appear Primitive?

31:56 Flock Safety and Urban Governance: AI Cameras, Crime Rates, and Local Choice

33:36 Las Vegas Case Study: Gunshot Detectors, Drones, and Real-Time Policing

34:46 Privacy by Design: Rolling Databases, Audit Logs, and the Boundaries of AI-Powered Public Safety

AI Layoffs and Tech CEOs’ Communication Failures

36:56 Cloudflare Layoffs and the “Measurers”: Why This Language Terrified Employees

37:30 Zuckerberg: Training Code Models on Internal Staff, Outsourced Workers Deemed Insufficient

38:52 Chamath Criticizes Tech CEOs’ PR: Don’t Reduce People to Labels

39:24 Shyam Sankar Clip: Listen to Factory Workers, Nurses, and End Users

41:11 AI Company Layoff Narratives: Training Models While Fearing Replacement

41:37 Episode Banter: Anthropic Hiring Jokes and Chamath’s Leg Gag

SpaceX’s $2 Trillion Imagination

43:55 SpaceX Files S-1: $1.75T Valuation, Targeting History’s Largest IPO

45:00 Three Core Businesses Breakdown: Starlink, Spaceflight, and xAI

45:32 Elon Web Services: Anthropic Renting Colossus Could Generate $10B+ Revenue

46:55 Gavin: SpaceX’s Speed and Cost Advantages in Building Data Centers

47:50 90-Day Termination Clause: Flexible Exit Mechanisms Between Anthropic and SpaceX

48:05 Cursor Composer 2.5: Proprietary Code Data + Colossus Compute Power Explosion

50:12 Will Elon Sell Compute to Google and OpenAI?

50:41 Grok Build: xAI Completes the Critical Agent Harness Puzzle

51:47 Harness Is as Crucial as the Model: Competition Over State, Memory, and Runtime

53:17 Friedberg: Space Communications and Space Data Centers as Civilization Backups

55:04 SpaceX Origin Story: From Biosphere Backup to Building Rockets In-House

56:16 Chamath: How the $2T Valuation Adds Up—Starlink, Colossus, and Execution Flywheel

60:51 DC-to-DC: Could Elon and Jensen Redefine Data Center Power Architecture?

61:39 Starship Scale: How Rapid Reusability Transforms Orbital Mass Economics

63:22 Gavin: The Huge Difference Between Standard and Rapid Reusability

65:24 Orbital Compute Timeline: H100s Already in Orbit; Commercialization Expected by 2028–2030

Nvidia, Chips, and AI Infrastructure

67:11 Nvidia Earnings Blow Up: Revenue, Profit, Cash Flow, and Buybacks All Hit Record Levels

69:31 Gavin: It’s Impossible That All AI-Related Company Valuations Are Correct

70:03 Misalignment in valuation between Nvidia, memory, power, cooling, and optical communication

71:10 ASIC narrative and Nvidia's market share: Why Jensen might feel frustrated

72:51 Shadow competition: Why TPU, Trainium, and Inferentia don't benchmark publicly

73:29 Nvidia's CPU business: What it means to reach $20 billion in a year

73:54 Chamath: Domain Specific Architecture is happening inside Nvidia

74:34 GPU financing cost and lifespan: Why depreciation cycles matter

75:43 After inference splitting, older GPUs may have a 10 to 15-year effective lifespan

76:25 GPU asset-backed financing: Neocloud's deep advantage

76:36 CoreWeave and Jensen: How Nvidia's earnings "saved" the new cloud company

Macroeconomic market: Debt, inflation, and U.S. advantages

79:22 Market update: Oil prices, inflation, bond yields, and global interest rate pressure

80:32 Dr. Doom moment for Friedberg: Global debt at 310% of GDP

81:20 High debt spiral: Printing money, inflation, asset prices, and credit crisis

82:04 Chamath: Hold only a few companies you truly believe in, avoid speculation

83:03 Gavin: Rising interest rates are concerning, but AI fundamentals are unprecedented

84:37 The U.S. is still "the best house in a bad neighborhood"

85:26 Strait of Hormuz: Why closing it is relatively advantageous for the U.S.

86:16 Energy self-sufficiency, natural gas, and reindustrialization: Structural advantages of the U.S.

86:56 Does AI have seasonality? Student summer breaks, agentic AI, and demand fluctuations

93:26 The chessboard of oil: Iran, Venezuela, Russia, and the Strait of Hormuz

93:39 War costs and global alliances: Why energy affects Taiwan Strait risks

93:46 Closing remarks: Sacks' absence, special thanks to Gavin Baker and his father-in-law Jeff Painter

🌟 Highlights

💡 Recursive self-improvement: A new Moore's Law for AI?

Karpathy's move to Anthropic was seen by several guests as a significant signal: AI might be entering a new phase of model self-improvement, self-experimentation, and self-training. Chamath believes that if recursive self-learning combines with massive computing power, model capabilities could improve by an order of magnitude each year, forming a kind of "new Moore's Law." Gavin further added that recursive self-improvement and continual learning might be the last two frontiers of AI.

"The idea of recursive self-learning would put these models into both hyperdrive and autopilot."

🧠 Why is the U.S. starting to fear AI?

The show spent a lot of time discussing the AI public relations crisis: young people booing AI at graduation ceremonies, tech company CEOs laying off workers en masse while claiming AI will replace jobs, and model founders exaggerating risks, leading to increasing public fear. Friedberg believes that AI triggers a deep-seated unease: a small number of people who control the technology will gain enormous leverage, while the majority have yet to see how they benefit.

"The only thing ordinary people are being told is that some people are making trillions of dollars."

🚀 SpaceX is more than just a rocket company; it's the future of internet infrastructure

SpaceX's S1 became a major focus of this episode. Starlink is already a high-growth, high-profit business; xAI is catching up with cutting-edge models; and so-called "Elon Web Services" could turn SpaceX into an AI computing infrastructure company. Gavin particularly emphasized that SpaceX builds data centers faster than other companies, and the $10 billion contract for Anthropic to rent Colossus might just be the beginning.

"GPUs will go to those who can plug them in, turn them on, and start converting electrons into tokens."

🛰️ Space data centers: A backup system for civilization

Friedberg views Starlink and orbital computing from a broader perspective: if ground-based internet and data centers are affected by governments, regulations, geopolitics, or even war, then a space-based communication and computing network could serve as a backup for civilization. Gavin provided a more specific timeline: orbital computing might begin to truly commercialize between the second half of 2028 and the first half of 2030.

"Having a space-based communication network and space-based data centers is generally a good thing. It's always good to have a backup."

🧩