The AI Cold War, Signalgate, CoreWeave IPO, Tariff Endgames, El Salvador Deportations
Speakers
GavinSBaker
— Portfolio Manager, Atreides Management
Chamath Palihapitiya
— CEO, Social Capital
Jason Calacanis
— Angel Investor / Founder, LAUNCH
David Friedberg
— CEO, The Production Board
Summary
Gavin Baker of Atreides Management joins Chamath Palihapitiya, David Friedberg and Jason Calacanis to discuss Nvidia's balance sheet, the Hopper to Blackwell product transition and the CoreWeave IPO. The conversation then turns to US-China AI competition and export controls, and to whether AI agents become the next driver of compute demand while hollowing out incumbent software economics. The second half covers the administration's tariff and tax endgame and the deficit math behind DOGE, then Signalgate and the El Salvador deportations.
- Gavin Baker defends Nvidia's rising accounts receivable as a product of the Hopper to Blackwell transition rather than round-tripping.
- Baker argues AI compute is not overbuilt: China is buying every GPU it can, DRAM prices keep rising and OpenAI is gating products for lack of capacity.
- He takes the contrarian side on the CoreWeave IPO, arguing that running very large GPU training clusters is an underappreciated operating skill.
- M&A and IPO activity is reopening, with Wiz, CoreWeave and the Weights and Biases deal cited as evidence of returning animal spirits.
- Export controls are seen as slowing China near term while creating strong incentives for a domestic Chinese chip ecosystem over a decade.
- Agents and the model context protocol are framed as the next compute demand driver and as a threat to incumbent SaaS economics.
- The group debates the tariff endgame, deregulation, the tax-cut offset and the deficit arithmetic facing DOGE.
- Signalgate, federal records law and the El Salvador deportations are discussed as governance and human-rights issues without a market expression.
Ideas
Nvidia receivables reflect Blackwell transition, not fraud.
Baker says the bear case on Nvidia's balance sheet is misread. Nvidia's investments in neoclouds such as CoreWeave did not manufacture revenue - those GPUs would have been sold to Meta, Amazon and Microsoft anyway - and the real purpose was to break up a three-player hyperscaler oligopoly and seed Nvidia's own reference architecture, the same playbook Intel ran with Intel Inside, leaving Nvidia with a more fragmented customer base that has less power over it. The jump in accounts receivable from roughly 1.5 billion to 5.5 billion reflects the largest product transition in semiconductor history, Hopper to Blackwell, where a rack goes from about 1,000 to 3,000 pounds, from 60kW to 120kW and from air to liquid cooling, so revenue is recognized while systems are still being stood up at customers. He would only treat receivables as a genuine warning sign if the trend persists past the July quarter.
CoreWeave cluster operations harder than consensus assumes.
With sentiment on X negative, the raise cut from 2 billion to 1.5 billion and consensus calling CoreWeave an undifferentiated commodity business with heavy debt, heavy capex and one dominant customer, Baker takes the other side. He argues that operating very large GPU training clusters is genuinely hard - synchronizing tens of thousands of GPUs that are melting, cables that get unplugged, training runs that are lost - and that very few operators on earth do it well, so the business may not be the commodity everyone assumes. He compares the skepticism to the early dismissal of AWS and of Microsoft's shift from capex-light software to capex-heavy cloud, notes that the NPV of CoreWeave's contracts and its asset value are knowable, and says the Weights and Biases acquisition further differentiates it. His conclusion is that it is a good company whose offsetting positives are being ignored.
Export controls seed China's own chip ecosystem.
Baker says US export controls put enough friction into the system to give America a theoretical advantage, but policing GPU smuggling is harder than stopping drugs, because a Blackwell is small and vastly more valuable per unit and shipments route through Singapore or Vietnam, so it stays a cat-and-mouse game until some grand bargain. The cost of the policy is that it creates an immense incentive for China to build its own semiconductor ecosystem and to be algorithmically innovative, which DeepSeek already demonstrated. He puts the odds that China produces something competitive with or better than Nvidia within five years at roughly zero, but says that over ten years who knows, since the CCP thinks in centuries while America thinks in decades.
Agents gut incumbent SaaS cost structures.
Palihapitiya identifies incumbent American software - what he calls the software industrial complex - as the most exposed pool of value. It is a three and a half to four trillion dollar SaaS industry growing about 10 percent a year, adding roughly 300 billion of enterprise value annually that many customers would say was never really delivered. Whether the attacker is a startup or China, the most economically disruptive move available is to use a three-person team to displace a 30,000-person one, and if agents scale, the opex of producing the same output falls by one and possibly two orders of magnitude. Incumbents built on the old cost structure cannot compete with that, which is why he frames this, rather than Chinese hardware, as the big risk.
Humanoid robots scale far slower than assumed.
Answering the argument that reshored factories will simply be staffed by Optimus or Figure robots, Baker pushes back on the timeline: even counting China and Tesla together, he thinks it will take a long time to manufacture humanoid robots in vast numbers. Automation is therefore a real long-run headwind to the tariff-and-reshoring plan, but not an imminent substitute for factory labor, and investors should not underwrite near-term mass deployment of humanoids.
Up Next
This All-In Podcast video, published March 29, 2025, features GavinSBaker, Chamath Palihapitiya discussing NVDA, CoreWeave, KSTR, IGV, Humanoid robots. 5 trade ideas extracted by AI with direction and confidence scoring.
Speakers: GavinSBaker, Chamath Palihapitiya · Tickers: NVDA, CoreWeave, KSTR, IGV, Humanoid robots