Google’s AI Brain Drain, SpaceX's Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI

The episode opens with discussion of significant upheaval at Google's AI division. Demis Hassabis has been promoted to chair of DeepMind and chief scientist, while Jeff Dean—a legendary engineer who joined Google in 1999—is leaving to start Discovery Loop, an AI startup focused on deep scientific breakthroughs. Several other top researchers have also departed, with reports indicating morale issues and Gemini 3.5 Pro being months behind schedule. Google's stock fell 4% on the news, representing approximately $200B in lost market cap.

The hosts debate why this brain drain is occurring. David Freeberg argues that Google's board has made a strategic choice to prioritize capex deployment in compute infrastructure (which offers high returns with lower risk) over frontier model development (which is high alpha but very high beta). This creates a fundamental misalignment with scientists who want to focus on cutting-edge AI development. Brad Gerson notes this creates channel conflict—Google Cloud wants compute to rent to competitors like Anthropic, while internal teams need it for their own models. Sax argues the market structure is shifting toward a frontier model duopoly: only Anthropic and OpenAI are truly competing at the frontier, while other companies either focus on infrastructure or get left behind.

The conversation shifts to SpaceX's first earnings report as a public company. Despite the stock dropping 13% post-earnings and down 30% since IPO, the financial results were spectacular: $7.8B in Q2 revenue (up 92% YoY), with AI revenue more than tripling to $2.6B quarter-over-quarter from Elon Web Services (renting Colossus compute to Anthropic and Google). Capex surged to $18.4B in the quarter (6x YoY). The hosts discuss Elon's guidance of $100B ARR by year-end and his goal to pull forward the $1 trillion ARR target from 2031 to 2030.

Multiple revenue streams justify the valuation: Starlink's connectivity business generated $2.6B in adjusted EBIT on $4.3B revenue with 12M subscribers (doubled YoY); the space/launch business broke even; and AI generated $1.1B. Brad emphasizes that Starlink alone could become a $1 trillion market cap business within 18 months if valued at 30x its eventual $30-40B in revenue and free cash flow. Sax notes the importance of Starship's successful test flight and upcoming V3 satellites, which will provide 10x more bandwidth per launch than current V2 satellites, enabling direct-to-cellular connectivity and higher network capacity.

The hosts express confidence in SpaceX's ability to execute on its ambitious capex plans (targeting 2 gigawatts of compute by year-end, scaling to 5-10GW in 2025). They acknowledge market concerns about spot pricing for compute (currently $30-50/watt), financing $300B in capex, and whether demand will sustain as open-source models improve. However, they believe frontier labs will continue paying premium prices for access to the best compute.

The conversation then addresses Airtable's $1.28B acquisition by Italian private equity firm Bending Spoons—a 90% decline from its $11.7B peak valuation in 2021. Despite being a profitable SaaS company with $480M ARR growing 20% annually, Airtable was acquired for only 2.5x revenue. Sax explains that Airtable's board pushed for salesforce-led growth to achieve venture-scale returns, which failed—only 30% of the sales team made quota. This created an opportunity for Bending Spoons, which can eliminate 80-90% of the cost structure, return to product-led growth, and likely generate $300-400M in annual EBIT while maintaining 10-20% growth.

The broader SaaS discussion includes debate about whether the entire category is being disrupted. Sax and others argue that while no-code tools like Airtable are vulnerable to AI agents (Claude can now do what required learning Airtable), enterprise mission-critical software (Salesforce, Microsoft, Workday) won't be replaced because of compliance, integration, and switching costs. The hosts note that high-growth software ETFs (IGV) have actually performed well recently, with companies like Snowflake up 88% in six months, suggesting selective rather than broad SaaS disruption.

The final major topic concerns U.S. data labeling startups (Scale AI, Miro, Surreal) selling valuable training data to both U.S. frontier labs (OpenAI, Anthropic) and Chinese AI companies (Tencent, Baidu, Alibaba, Moonshot). Forbes reports China's top six labs spend $500M annually on this data. Jason argues this is helping China catch up and isn't patriotic; the data represents expert-created synthetic datasets with verified technical content that took PhDs to produce. Sax counters that data is largely a commodity, China has abundant talent that can recreate datasets, and banning sales risks trade wars and reciprocal restrictions on U.S. companies. Brad notes that while the U.S. is currently winning in AI, these stories create friction in Washington and will get more scrutiny if America's lead diminishes. The hosts ultimately agree targeted strategic controls make sense but dispute whether data labeling meets that bar.