SpaceX's $2T Case, Nvidia's Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?
Episode
102 min
Read time
3 min
Topics
Artificial Intelligence, Science & Discovery
AI-Generated Summary
Key Takeaways
- ✓Recursive AI Self-Improvement: Anthropic's hiring of Andrej Karpathy to lead pre-training signals a strategic bet that AI-driven model improvement will outpace human-led development. Karpathy's prior work—including auto-research, an open-source training tool that earned 82,000 GitHub stars in days—demonstrates his ability to architect systems where models run five-minute self-improvement experiments autonomously. Combined with Anthropic's existing infrastructure, this approach could compress the timeline to 10x annual model quality gains.
- ✓SpaceX Elon Web Services Revenue: SpaceX's S-1 reveals Anthropic pays $1.25B per month—$15B annually—to rent Colossus compute clusters, effectively adding a second Starlink-scale revenue stream overnight. SpaceX builds data centers in 66–122 days, faster than any competitor. With $20B in 2024 CapEx (60% toward AI compute), and the ability to stamp out gigawatt-scale facilities at declining timelines, SpaceX's AI infrastructure business could quadruple within two years as additional offtake partners emerge.
- ✓AI Sector Cross-Valuation Inefficiency: Gavin Baker identifies a structural mispricing across AI-related equities: memory makers trade at 3–5x PE, Nvidia trades at low-to-mid teens forward earnings, while power, cooling, and optical names are priced for dramatically higher growth. Both sets of valuations cannot simultaneously be correct. Investors should identify which segment's implied growth assumptions are realistic and position accordingly, rather than treating the AI infrastructure trade as a monolithic theme.
- ✓Cursor Composer 2.5 Benchmark Signal: After just three to four weeks of reinforcement learning on Colossus 2 using Cursor's proprietary coding dataset—reportedly larger than all public coding data on the internet—Cursor's Composer 2.5 model became Pareto dominant across all frontier coding benchmarks. This result used the same base model as Composer 2, demonstrating that proprietary training data plus concentrated compute access can leapfrog competitors without a new base model, a replicable playbook for any lab with unique token datasets.
- ✓GPU Useful Life and Financing Implications: The bear thesis that GPU amortization schedules of four to six years overstate AI infrastructure profits has been invalidated. CoreWeave secures asset-backed financing at approximately 6% interest with customers signing six-year contracts and paying in advance. Nvidia's Q1 2025 results, showing CPUs becoming a $20B annual business overnight, confirm that domain-specific architectures are consolidating inside Nvidia rather than fragmenting to ASICs, extending GPU useful life and strengthening the financing case for neo-cloud operators.
What It Covers
Chamath, Jason, Friedberg, and guest Gavin Baker from Atreides Management cover Andrej Karpathy joining Anthropic to lead recursive self-improvement research, SpaceX's $1.75T IPO filing revealing a $15B annual Anthropic compute deal, Nvidia's $81.6B quarter, AI's public relations crisis, rising global bond yields, and the geopolitical implications of the Trump-Xi summit.
Key Questions Answered
- •Recursive AI Self-Improvement: Anthropic's hiring of Andrej Karpathy to lead pre-training signals a strategic bet that AI-driven model improvement will outpace human-led development. Karpathy's prior work—including auto-research, an open-source training tool that earned 82,000 GitHub stars in days—demonstrates his ability to architect systems where models run five-minute self-improvement experiments autonomously. Combined with Anthropic's existing infrastructure, this approach could compress the timeline to 10x annual model quality gains.
- •SpaceX Elon Web Services Revenue: SpaceX's S-1 reveals Anthropic pays $1.25B per month—$15B annually—to rent Colossus compute clusters, effectively adding a second Starlink-scale revenue stream overnight. SpaceX builds data centers in 66–122 days, faster than any competitor. With $20B in 2024 CapEx (60% toward AI compute), and the ability to stamp out gigawatt-scale facilities at declining timelines, SpaceX's AI infrastructure business could quadruple within two years as additional offtake partners emerge.
- •AI Sector Cross-Valuation Inefficiency: Gavin Baker identifies a structural mispricing across AI-related equities: memory makers trade at 3–5x PE, Nvidia trades at low-to-mid teens forward earnings, while power, cooling, and optical names are priced for dramatically higher growth. Both sets of valuations cannot simultaneously be correct. Investors should identify which segment's implied growth assumptions are realistic and position accordingly, rather than treating the AI infrastructure trade as a monolithic theme.
- •Cursor Composer 2.5 Benchmark Signal: After just three to four weeks of reinforcement learning on Colossus 2 using Cursor's proprietary coding dataset—reportedly larger than all public coding data on the internet—Cursor's Composer 2.5 model became Pareto dominant across all frontier coding benchmarks. This result used the same base model as Composer 2, demonstrating that proprietary training data plus concentrated compute access can leapfrog competitors without a new base model, a replicable playbook for any lab with unique token datasets.
- •GPU Useful Life and Financing Implications: The bear thesis that GPU amortization schedules of four to six years overstate AI infrastructure profits has been invalidated. CoreWeave secures asset-backed financing at approximately 6% interest with customers signing six-year contracts and paying in advance. Nvidia's Q1 2025 results, showing CPUs becoming a $20B annual business overnight, confirm that domain-specific architectures are consolidating inside Nvidia rather than fragmenting to ASICs, extending GPU useful life and strengthening the financing case for neo-cloud operators.
- •AI Public Perception and Messaging Failures: Tech CEO communications are accelerating anti-AI sentiment. Cloudflare's CEO publicly labeled laid-off employees "measurers" while reporting record revenue, and Meta simultaneously announced 8,000 layoffs while installing recording software on all employee computers to train models. Both actions confirm the public's fear that AI adoption means workers train their replacements. The actionable counter-strategy: redirect media narratives toward verified end-user outcomes—specific drug discoveries, crime reduction data, productivity gains—rather than model capability announcements.
- •Global Bond Market Stress Indicators: Japan's 30-year yield hit a record 5.1%, UK yields reached their highest since the 2008 financial crisis, Germany's highest since 2011, and Polymarket prices a 99% probability that May US CPI prints at 4.2% or higher, with professional forecasters projecting CPI reaching 6% in Q2. The Federal Reserve rate-cut narrative has reversed toward potential hikes. Gavin Baker argues America remains relatively insulated due to energy self-sufficiency, natural gas-based electricity generation, and dominant AI and public market positions versus all other major economies.
Notable Moment
Gavin Baker described a hedge fund manager whose daughter was born with a rare neurological condition causing severely reduced neural firing. Using LLMs to research existing approved drugs, the father identified a treatment that raised her neural activity from 30–40% to 80–90%. Baker added the father now believes AI-assisted drug tailoring could produce a complete cure within months.
You just read a 3-minute summary of a 99-minute episode.
Get All-In with Chamath, Jason, Sacks & Friedberg summarized like this every Monday — plus up to 2 more podcasts, free.
Pick Your Podcasts — FreeKeep Reading
More from All-In with Chamath, Jason, Sacks & Friedberg
Trump-Xi Summit, Benioff: "Not My First SaaSpocalypse," OpenAI vs Apple, Multi-Sensory AI, El Niño
May 15 · 76 min
Marketing School
The AI Search Strategy That Actually Works
May 25
More from All-In with Chamath, Jason, Sacks & Friedberg
Charles & Chase Koch on How They Quietly Built a $150B Empire
May 12 · 95 min
a16z Podcast
Why AI Isn’t Killing SaaS Yet
May 25
More from All-In with Chamath, Jason, Sacks & Friedberg
We summarize every new episode. Want them in your inbox?
Trump-Xi Summit, Benioff: "Not My First SaaSpocalypse," OpenAI vs Apple, Multi-Sensory AI, El Niño
Charles & Chase Koch on How They Quietly Built a $150B Empire
Spencer Pratt on Fixing LA: Wildfires, Homelessness, Corruption & the Fight to Take It Back
Elon's Anthropic Deal, The Next AI Monopoly?, "FDA for AI" Panic, Trading the AI Boom
OpenAI Misses Targets, Codex vs Claude, Elon vs Sam Trial, Big Hyperscaler Beats, Peptide Craze
Similar Episodes
Related episodes from other podcasts
Marketing School
May 25
The AI Search Strategy That Actually Works
a16z Podcast
May 25
Why AI Isn’t Killing SaaS Yet
Animal Spirits
May 25
Talk Your Book: Investing in the Rise of the Robots
Capital Allocators
May 25
Fundraising Mastery: The Tao of Kimmer – John Kim (EP.503)
The Productivity Show
May 25
The Productivity Stack: Apps and Tools We Actually Use Every Day (TPS614)
Explore Related Topics
This podcast is featured in Best Tech Podcasts (2026) — ranked and reviewed with AI summaries.
Read this week's AI & Machine Learning Podcast Insights — cross-podcast analysis updated weekly.
You're clearly into All-In with Chamath, Jason, Sacks & Friedberg.
Every Monday, we deliver AI summaries of the latest episodes from All-In with Chamath, Jason, Sacks & Friedberg and 192+ other podcasts. Free for up to 3 shows.
Start My Monday DigestNo credit card · Unsubscribe anytime