Google AI Exodus, SpaceX Earnings, Airtable's Collapse

Chapters

1. 0:00 AI and tech

   Google's AI brain drain and capital shift

   The hosts open with a Chamath road-trip photo and then dig into Google's AI shakeup: Demis Hassabis moves to chair of DeepMind and chief scientist while Jeff Dean and three researchers leave to found Discovery Loop. Google shares fall 4%, wiping out about $200B in market cap. Friedberg argues Google is rationally shifting capital to $200B in AI infrastructure capex, which is tax-advantaged and low-beta, rather than frontier model development, which is high-beta and faces open-source competition. Gerstner adds that Google Cloud's incentive to rent compute to Anthropic creates an internal channel conflict that pushes top scientists out.

2. 9:20 AI and tech

   Who wins the frontier-model race

   Sacks reads the Polymarket odds for the top AI model by Dec 31: OpenAI 32%, Google 20%, Alibaba 14%, others ~10%. He argues frontier intelligence is a two-tier market, with Anthropic and OpenAI a premium duopoly (Anthropic ARR likely $110–120B exit), while non-frontier open-weight models are commoditized and charge only for compute. Jason counters that open-source models are good enough for 95% of his use cases; Elon disagrees, saying frontier models are far ahead. Friedberg sees a blend: cheap open-weight for simple workflows, premium specialized models for life sciences and video, with Gemini strongest in verticalized domains.

3. 20:41 Markets

   SpaceX's huge quarter and Starlink

   SpaceX reported its first public quarter: $7.8B revenue (+92% YoY, +67% QoQ), Elon Web Services compute rental tripled to $2.6B, and capex hit $18.4B. The stock fell 13% after IPO, settling at a $1.4T valuation. Brad calls it a solid beat, noting guidance of $100B ARR by year-end and pulling the $1T target forward to 2030. Sax highlights Starlink's $2.6B adjusted EBIT on $4.3B revenue, 12M subscribers doubling YoY, and Starship/V3 satellites enabling huge bandwidth growth. They debate GPU spot pricing ($30–$50 per watt), financing $300B for 6GW of new compute, and Elon's hardware execution advantage. The segment closes with an All-In Summit promo.

4. 47:42 Deals and companies

   Airtable's sale and the SAS reckoning

   Airtable, a profitable SaaS company with ~$480M revenue growing 20%, was acquired by Bending Spoons for $1.28B, about 10% of its 2021 peak valuation; with cash, the deal was $2.25B. Sacks explains Airtable spun out its AI agent business Hyper Agent, while Bending Spoons will likely cut 85–90% of costs and return to product-led growth, potentially generating $300–400M EBITA. Brad calls it a classic venture-to-PE transition made easier by AI learning codebases. Friedberg says the ZIRP-era and AI capex paradigms are fundamentally different: SaaS overvaluation came from treating companies as guaranteed annuities, while AI infrastructure is capital allocation, not multiple arbitrage. They caution against extrapolating to all SaaS, citing Microsoft compliance rails and Salesforce adoption by government.

5. 1:05:55 Policy and regulation

   US training data fuels China's AI

   Forbes reports US data-labeling startups (Surge AI, Mercor) sell PhD-written training data to Chinese labs like Tencent, Alibaba, and Moonshot, spending roughly $500M a year. Sacks argues data is largely a commodity, China has plenty of PhDs, and banning sales could invite retaliation; he supports targeted strategic controls like EUV export restrictions. Gerstner agrees but warns Washington will scrutinize the issue if America loses the frontier lead. Jason counters that these expert-verified datasets are the secret sauce behind LLM improvements, and selling them to China hands over a US advantage. He also notes immigration policies push top graduates out, and the show wraps with sign-off.