My biggest takeaways from this week's All-In, with Marc Benioff guest-hosting:

1. The surest path to no conflict with China is economic entanglement. Chamath reads the Trump-Xi summit as an entanglement play: bring the CEOs, sell the planes, chips, soybeans and payment rails, and make war too expensive for both sides. He thinks the public theater hides the real work, which is two powers quietly settling into a bipolar carve-up of the world. Friedberg frames the same bet as the Thucydides question: share in a bigger pie, or fight over a static one. [5:53]

2. Withholding the top chips from China is irrelevant: their models are already as good without them. Marc says the premise of withholding Nvidia's best silicon has already failed, because Chinese labs learned to build competitive models on lesser hardware. His conclusion is to sell whatever the US can, on the theory that more commerce buys more peace. Chamath lands in the same place for a different reason: sell the chips so Nvidia wins and Huawei never gets the oxygen to emerge as a rival. [20:16]

3. Taiwan has 18 months left as the world's most important conversation. Chamath puts a clock on it: as US fabs get within a node or two of what Taiwan does, the island stops being the strategic chokepoint it is today. Friedberg points to the TSMC facility in Arizona already printing chips and asks whether Taiwan really matters once both sides own their own fabs. Marc goes further and calls the whole thing a nonsense issue that ends in China and Taiwan reconciling. [24:34]

4. There's a hypnosis around AI, and it hasn't shown up in the numbers yet. Marc calls the software crash the SaaSpocalypse and says it is not his first: Salesforce is down 37%, ServiceNow 42%, Workday 45%, and HubSpot trades at two times sales after a great quarter. His argument is that the market is pricing a fear, not a result, and until AI displacement shows up in revenue the re-rating is cyclical. Salesforce will do over $46 billion in revenue and more than $16 billion in cash flow this year, and he tells employees you can't get drunk on the stock price. [32:59]

5. The low end of software is finished. The high end is quite safe. Chamath splits the market in two rather than defending it whole. The low end has no safe space left; the large monoliths with deep C-suite relationships, where Marc operates, are positioned to win. His evidence is what he sees at the high end, where all the action is: the AI labs themselves are discovering that this is a lot harder than they thought. [36:01]

6. The next trade is the market asking: $3 trillion in four years, what is the ROI of these tokens? Chamath predicts the breathless phase ends with one question from public investors, and the AI labs will not be able to answer it alone. His read of OpenAI's $4 billion deployment deal, with a 17.5% guaranteed preferred return, is that selling into the enterprise is much harder than assumed. So the labs end up routing their tokens through incumbents like Salesforce to get them sold. [36:01]

7. Anthropic was right: bet on coding agents while everyone else chased sex bots, video and ads. Marc credits Anthropic's takeoff to one correct strategic bet while rivals scattered across consumer distractions. He speaks as a customer, not an analyst: Salesforce will spend roughly $300 million on Anthropic this year and he calls the coding agents a rocket ship that will not stop. That, in his telling, is what changes what Salesforce can build and sell. [49:07]

8. AI is probabilistic, so it only works when grounded in a single source of truth. Marc's case for the roughly $8 to 9 billion Informatica acquisition is that a model can kind of figure things out, but only when it is grounded in real data with a semantic layer underneath. Harmonized, federated customer data is that grounding, and the vendor who owns it owns the agent. Chamath pushes on where that semantic understanding actually lives, and for many companies the answer is Slack and email. [40:17]

9. Most tokens are wasted. A hot new company will sit between you and the frontier labs. Jason argues always-on multi-sensory AI, watching screen, audio and webcam for eight hours a day, will 1000x token demand. Marc says that is a mistake: most of those tokens do not need a frontier model, and a routing layer will send only the ones that do to Anthropic and OpenAI. He predicts the company that builds that layer is the next hot one, which means the labs never see most of the volume. [1:00:20]

10. A record super El Niño makes next year the hottest on record, and the risk is food, not weather. Friedberg sizes the stored ocean heat at about 11 million terawatt-hours against 25,000 terawatt-hours of annual global energy use, which he calls 500 years worth of human energy. Released into the atmosphere, that drives crop failures in Brazil, Australia and India. The scenario he worries about is 150 million Indian farmers facing a failed monsoon and short fertilizer, with a calorie deficit and economic crisis across South Asia over the next year. [1:03:01]

11. Layered SPVs need a bullet put in their head. Chamath applauds Anthropic for negating multi-layered SPVs sold to retail investors and wants every mechanism that lets companies stay private longer killed off. His preferred fix is structural: go public sooner and rationalize the cap table. His prediction is that once SpaceX, Anthropic and OpenAI list, a litany of lawsuits follows between the purveyors of those SPVs. [1:12:07]

12. Your valuation is fantasy land until somebody actually pays you. Marc's advice to private-company CEOs is to stop steering by the paper number and steer by customers who pay. It is the same discipline he applies at Salesforce in a drawdown: anchor on customer success and the long-term quality of what you deliver, because you can't get drunk on the stock price. [45:00]

— All-In Podcast, "Trump-Xi Summit, Benioff: "Not My First SaaSpocalypse," OpenAI vs Apple, Multi-Sensory AI, El Niño" (2026-05-15)