SpaceX's $2T Case, Nvidia's Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?
The All-In Podcast hosts Jason Calacanis, Chamath Palihapitiya, and David Friedberg are joined by Gavin Baker of Atreides Management to discuss major tech developments including Andrej Karpathy joining Anthropic, SpaceX's historic IPO filing at a $1.75 trillion valuation, NVIDIA's record-breaking earnings, AI's public relations challenges, and macro concerns around rising bond yields and inflation.
Summary
The episode opens with discussion of Andrej Karpathy joining Anthropic to lead a new pre-training team focused on recursive self-improvement. The hosts and Gavin Baker frame Karpathy as a generational talent who has been at the frontier of every major AI wave, from Tesla's FSD to co-founding OpenAI. Baker argues that recursive self-improvement and continual learning are the two remaining 'holy grails' of AI, and that unlocking them could put model quality on a parabolic improvement curve. Friedberg adds that re-architecting models — particularly networks of smaller specialized models — could dramatically reduce cost per token.
The group then addresses the AI public relations crisis, noting that commencement speeches by tech executives have been booed and that there is growing public hostility toward AI. Friedberg offers a multi-layered explanation: economic asymmetry where a small group captures early gains, possible foreign-state-funded anti-AI sentiment campaigns, and a deep psychological discomfort with AI's anti-humanist implications. Chamath argues that industry leaders are 'trading their own book' — particularly Dario Amodei, whom he accuses of using safety rhetoric to build a regulatory moat. The group agrees the solution is to spotlight end-user stories of AI benefit, such as a hedge fund manager using LLMs to find a treatment for his daughter's rare genetic condition.
The conversation turns to SpaceX's S-1 filing, targeting a $1.75 trillion valuation and a potential $75 billion IPO — the largest ever. Key revelations include Starlink generating $11.4 billion in revenue at 50% growth, the AI/compute business doubling year-over-year, and a bombshell: Anthropic is paying SpaceX $1.25 billion per month — a $45 billion, three-year deal — to rent Colossus 1 and 2. Baker highlights that SpaceX builds data centers dramatically faster than competitors, with successive builds taking 122, 91, and 66 days respectively. The group also discusses Cursor's Composer 2.5 model, which after just 3-4 weeks of reinforcement learning on Colossus became Pareto-dominant on coding benchmarks, and the new GrokBuild harness that gives XAI a competitive agentic runtime. Chamath argues that SpaceX's true value is in being the only entity capable of building gigawatt-scale data centers and that at 20x revenue it may look cheap in hindsight, especially if Tesla and SpaceX eventually merge.
NVIDIA's Q1 earnings are reviewed: $81.6 billion in revenue up 85% YoY, $58 billion net income, $48 billion free cash flow, 75% gross margins, $80 billion in new buybacks, and a surprise $20 billion CPU business. Baker argues the 'NVIDIA losing share to ASICs' narrative is incorrect — NVIDIA is growing faster than hyperscaler CapEx even without China revenues — and that rival ASIC chips are not being submitted to benchmarks like MLPerf because they would lose. He also notes the cross-sectional inefficiency in AI valuations: power/cooling/optical valuations and semiconductor valuations cannot simultaneously be correct.
On macro, the group notes 10-year Treasury yields at 4.6%, Japan's 30-year at an all-time high of 5.1%, UK yields at GFC highs, Germany at 2011 highs, and Polymarket projecting May CPI at 4.2%+. Friedberg invokes global debt-to-GDP at 310% and warns of a potential credit cascade. Baker pushes back, arguing America is uniquely advantaged: energy self-sufficient, largest oil/gas exporter, home to the best public and private AI companies, and with NVIDIA's fundamentals strengthening. He notes that every day the Strait of Hormuz is closed is relatively better for America than its competitors.
Finally, the hosts debrief Trump's visit to China, concluding it was productive but produced no grand deal — just some aircraft purchases, agricultural commitments, and possible chip sales to Baidu. Chamath speculates the real value was in private negotiations over geopolitical 'chessboard' arrangements. Baker argues selling deprecated NVIDIA GPUs to China actually reduces the likelihood China builds a rival AI ecosystem, which would be strategically worse for the US.
Key Insights
- Gavin Baker argues that Anthropic being EBITDA-positive, combined with the broader LLM market approaching $200-400B ARR by year-end, proves the ROI on AI infrastructure spend is real and not circular — directly countering the bear case on hyperscaler CapEx.
- Baker claims NVIDIA is actually gaining share against ASICs when you properly strip out China revenues and compare apples-to-apples with Broadcom's AI semiconductor growth, calling the 'NVIDIA losing share' narrative largely unsupported by submitted benchmarks.
- Chamath argues that Dario Amodei's safety-focused public messaging is a calculated business strategy to build a regulatory moat, noting that the volume of AI existential risk rhetoric has correlated suspiciously closely with successive funding rounds.
- Friedberg contends that anti-AI sentiment has three distinct drivers: economic asymmetry favoring a small group, possible foreign state-sponsored disinformation campaigns to slow US technological progress, and a deep psychological anti-humanist reaction similar to the Copernican revolution.
- Baker highlights that SpaceX built successive data centers in 122, 91, and then 66 days — dramatically faster than any competitor — making it uniquely positioned to scale the Elon Web Services compute rental business rapidly with Anthropic as the anchor tenant.
- The Cursor Composer 2.5 benchmark result — becoming Pareto-dominant over all other coding models after just 3-4 weeks of reinforcement learning on Colossus using Cursor's proprietary coding data — is cited as proof that compute access plus proprietary training data can produce rapid, dramatic model quality jumps.
- Baker argues that the divergent valuations across the AI supply chain (power/cooling/optical at high multiples vs. semiconductors at low multiples) represent a cross-sectional market inefficiency — both sets of valuations cannot simultaneously be accurate.
- Chamath frames SpaceX's long-term value around being the only entity realistically capable of building gigawatt-scale data centers, arguing the terrestrial compute rental business alone justifies a significant portion of the $1.75T valuation even before space-based compute.
- Baker argues that selling deprecated NVIDIA GPUs to China strategically reduces the probability that China builds an entirely separate AI hardware ecosystem, which would be more energy-inefficient and harder for the US to influence or monitor.
- Friedberg argues that a space-based communication and compute network — immune to government control or shutdown — represents a civilizational backup system, analogous to the strategic importance of nuclear deterrence, making SpaceX's orbital ambitions geopolitically significant beyond commercial returns.
- Baker contends that America is uniquely advantaged in the current global instability because it is both energy self-sufficient and the world's largest oil and gas exporter, meaning every day the Strait of Hormuz is closed disproportionately harms China, Europe, Japan, and India relative to the US.
- Chamath argues that NVIDIA's $20 billion CPU business announced this quarter — making them overnight one of the world's largest CPU manufacturers — demonstrates that NVIDIA is successfully executing domain-specific architectures by co-designing chips in partnership with every major AI lab simultaneously.
Topics
Full transcript available for MurmurCast members
Sign Up to Access