AI Sovereignty Wars, Palantir-Nvidia Deal, SCOTUS Birthright Ruling, Newsom's CA Budget Lie

Episode 279 opens with the hosts announcing what they call a "rocket docket," and the lead item is a partnership between Palantir and NVIDIA to build what Palantir is branding a sovereign AI operating system. The framing question Jason Calacanis poses is deliberately loaded: sovereign AI, he notes, is a term the panel has used before. Palantir's contribution is not a model of its own but a wrapper — it will use NVIDIA's Nemotron open models to construct a custom, frontier-quality model aimed at US government customers, with agencies owning the hardware, the data, and the model weights.

The manifesto Palantir pushed alongside the announcement is the part worth lingering on, because it states a thesis in two sentences: "data retention is your treasure," and transferring it hands over both your existing winning plays and the means of production for new ones. That is a commercial argument dressed as a security argument, and the rest of the segment is essentially the panel testing whether the commercial argument holds up.

The clip they play is Alex Karp on CNBC, which the hosts describe as a Robin Williams-style monologue and which some coverage apparently called a televised nervous breakdown. The panel's own label is friendlier — "Alex Karp on a Tuesday" — but the substance of the clip is a direct attack on the frontier labs, delivered in Karp's usual register of grievance and contempt.

Karp's clip is short and quotable, and two lines carry most of the weight. The first is about who controls the weights when the government comes asking: "When the d o Department of War goes to you and says, I need this application, do they get to control the weights to do it, or do you get to control the weights?" The second escalates that into a political question — "Are we really gonna outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane."

David Sacks, who spent the first half of the administration as AI czar, pushes back hard on the framing of the clip as a meltdown. "It was nothing of the sort," he says, arguing that the crash-out label came from legacy media and is itself evidence that Karp said something worth hearing. Sacks isolates the claim he had not previously considered in those terms: that what technical customers want is "control over their compute, their models, their data stack, and their alpha" — alpha meaning proprietary knowledge — and that they want to know "they own the means of production" and that it is not being transferred to someone else.

This is where the episode's real argument starts, and it is narrower than the sovereign-AI branding suggests. Sacks is not defending Palantir or NVIDIA; he is translating Karp's complaint into a claim about enterprise risk: a company that hands its know-how, trade secrets, and customer data to a model provider is exposed to a counterparty that may eventually decide to compete with it. Framed that way, AI safety for an enterprise stops being about model behavior and becomes about control of the stack.

Karp's own language in the clip is more scattered than Sacks's reconstruction. He complains that the basic view among enterprises is "I'm gonna chillax, and waste my time with tokens," that they get no value while the vendor gets their IP, and he concedes that "there's nothing more fun than debating Dario in private" before insisting he is not throwing shade. The panel lets the messiness stand; nobody tries to smooth it into a cleaner thesis.

Sacks then supplies the empirical spine of the segment. The example is Figma. According to The Information, Anthropic "blindsided its then business partner" with the launch of Claude Design, a vertical app competing directly in Figma's category. The detail that lands hardest is governance: Anthropic's chief product officer had served on Figma's board and did not resign until three days before launch. Figma's stock, Sacks says, has fallen roughly 50% this year while Anthropic's valuation surged.

He treats this as a pattern rather than an incident, listing Claude Science, Claude Security, Claude Legal, Claude Financial, and Claude Code as vertical expansions into categories previously served by companies building on Anthropic's own models. The causal story he tells about Claude Code is the sharpest part: Cursor, one of Anthropic's largest customers, created the coding-assistant category first; Anthropic watched where value was being created on top of its models and moved in directly. "Why don't we vertically integrate?" is how he renders the internal logic.

The analogy he reaches for is Microsoft and Google — dominate the model layer as if it were an operating system, then use that position to capture the most lucrative verticals. Microsoft had the Windows monopoly and then took spreadsheets, word processing, and eventually the browser; Google had search and, over time, stopped sending users off-site, to the point where Sacks says fewer than half of searches now leave Google properties. The analogy is suggestive rather than proven — the panel never tests how far the comparison actually extends — but it does the work of making Karp's complaint sound structural instead of personal.

From there Sacks turns prescriptive and blunt: an enterprise or developer sharing proprietary data with such a provider is "mortgaging your future" and "sealing your fate." The line is rhetoric, not analysis, but it sets the tone for everything the other three hosts say afterward.

The most interesting move in the segment is Sacks inverting the safety argument. Dario Amodei, he notes, argues simultaneously that open models are dangerous and should be restricted, and that his company should expand into adjacent verticals. Sacks's question is "dangerous to whom?" — not to enterprises that want to retain control of their data, he answers, but to a business model that requires customers to have little choice at the model layer.

His conclusion is that genuine enterprise AI safety requires retaining the ability to choose who sees your alpha. Whether or not one accepts the definition, the reframing is the episode's cleanest piece of argument: it takes a term — safety — that the frontier labs have used to justify restriction, and redefines it as a property of customer-side control rather than model-side guardrails. Calacanis's only response is to note that Sacks has picked "that carcass to the bone," which is a fair description of the segment's length.

Chamath Palihapitiya answers with data rather than theory, organized as three posts. The first is a BCG study on return on capital employed. With long-term rates where they are, the cost of capital has returned to its long-run average of roughly 8% to 11%, and the problem is that half of large US companies cannot clear that bar. He adds a second layer: persistently low returns in the 1% to 3% range affect roughly 45% of companies worldwide — he glosses this as "about one in seven," an inconsistency nobody on the panel flags.

The point of the detour is to establish that most businesses are already marginal, which makes the magic-box offer dangerous: a vendor arrives promising that if you tell it everything you do, it will make everything better, and then "from the shadows, the magic box says, you know what? I've decided to compete with you." Given how many firms are already failing to earn their cost of capital, Chamath argues, paying for the privilege of ending up with the same answer as your competitor is irrational.