The episode opens with Brad Gerstner subbing in for David Friedberg on his vacation, Jason Calacanis broadcasting from Paris after conducting eight interviews at the RAISE conference, and Chamath Palihapitiya calling in from what he calls a 'hot software summer' of enterprise sales. Chamath recounts a dinner in Geneva hosted by Marc Benioff, attended by Jensen Huang, Brad Smith, and Anthony Tan of Grab, connected to a UN Commission on AI for which Benioff is co-chairman. He gives Benioff an effusive compliment — 'this guy is a fucking master' — as a window into what it means to build an enterprise sales machine at global scale. Brad, fresh from a week in Washington DC, drops hints about what's coming in the Trump Accounts segment. The brief intro sets up the episode's dual registers: hardheaded investor analysis and genuine enthusiasm for a specific policy mission.

Jason brings in real-world evidence: Uber's CTO publicly posted that 99% of Uber engineers use AI tools, over 70% of pull requests are attributed to AI agents, and the company has built 2,500 agentic skills. Uber's answer to the ROI question is forward-deployed engineers embedded in operational departments to identify and build agentic pods. DoorDash's CTO Andy Fang goes further, releasing internal benchmarks showing they've introduced open-weight models into their AI code review workflow without degrading quality — using frontier Fable for hard tasks and Kimi 2.6 for lower-level work. Jason frames this as the 'tip of the spear' behavior that will spread across enterprise. Brad contextualizes it: yes, there's optimization happening, but revenue growth is not slowing down because every new use case being discovered still wants the most powerful frontier model available. His analogy: Snowflake optimized heavily under the hood throughout its growth phase, but revenue never slowed. He caps it with a jaw-dropping claim — if these labs end the year at $100 billion, they could 3–5x again the following year, adding $200 billion of incremental revenue, an amount that has no historical precedent.

David Sacks synthesizes the discussion with a framework: enterprises want model fungibility — the ability to hot-swap models for the cheapest one that gets the task done. But there's a hard technical blocker: memory, context, and history are all tied to the model, and nobody has figured out how to make them fully portable. So even technically capable teams like Coinbase and DoorDash can build routing middleware, most enterprises simply cannot. The data bears this out: open source went from 19% to 11% of enterprise AI wallet share year-over-year. He introduces the Decagon framework — open models are great for mature, well-defined use cases where you can post-train on specific data. But for immature use cases, which is everything enterprises are still figuring out, you want the most capable general intelligence available. Jason brings in Ali Ghodsi's Databricks finding: the same GLM 5.2 model, but with a different harness, produced 2x token savings. This is profound — it means much of the optimization opportunity has nothing to do with model choice and everything to do with how you structure the call. Jason shares his own optimization experience: asking his agents to self-optimize their token usage reduced consumption by 80%. The session ends with the group agreeing that the tip of the spear — the 1% of technically capable deployers — are working it all out, and the rest will follow.

Jason pivots to the week's geopolitical AI story: Reuters, citing anonymous sources, reported that Chinese regulators held meetings with Alibaba, ByteDance, and ZhipuAI to discuss limiting overseas access to China's top AI models — both open and closed. The CCP is simultaneously making AI research leaks a national security offense and tightening control over who can fund Chinese AI labs. Manus, a Chinese agentic AI startup, had employees pulled back from Singapore by the CCP. Sacks walks through why this makes sense: Chinese labs like Alibaba's Qwen and ZhipuAI's GLM went open specifically because they were behind the frontier and needed developer adoption and reinforcement learning from usage. Now that they're catching up, the economic logic shifts toward closing. He draws the explicit parallel to Sam Altman, who took OpenAI from open to closed in exactly the same move. Brad adds a provocative claim: GLM 5.2 contains watermarks from Anthropic's models, suggesting distillation — effectively downloading frontier intelligence built with American capital. He expects the US government to take steps against distillation. Brad also reports from his DC meetings that from the president down, the single unifying principle in Washington is doing everything necessary to stay ahead of China in AI.

Sacks pivots from the philanthropy narrative to what he thinks is an underappreciated story: Trump Accounts are among the most powerful tax-advantaged vehicles ever created in America. He walks through the full stack: donate up to $5,000 per year per child, with family, friends, and employers all eligible to contribute. The account compounds tax-free until 18. Employers can contribute $2,500 tax-free to an employee's child's account — a no-brainer for both parties. At 18, the child can roll it into an IRA or, better yet, a Roth IRA. The optimal move, per the CPAs Sacks has been reading: wait until the child is no longer a dependent and is in the 0% tax bracket — probably in college or just after graduation — and do the Roth IRA conversion at minimal cost. A $200–300K account at age 18 compounds to over $10 million by age 60 at historical market rates. The account also allows penalty-free withdrawals for home down payments, health emergencies, college, and business startup. Maxed out with contributions from friends, family, and employer, Sacks says Sacks, if a Trump account had been maxed out using historical market returns, the child would be a millionaire by age 28. He summarizes: this is like getting an IRA at birth, which is something that has never before been possible in America.