Gavin Baker: The AI Cold War, Signalgate, CoreWeave IPO, Tariff Endgames, El Salvador Deportations

The episode opens with banter about made-up insults and Friedberg's dog, Marshall, who is accused of jumping on the table, eating the nuts and almost ruining a Christmas dinner. Another participant's dogs are shown by contrast as well-behaved, and their breeder is named as Valentina Zanini.

The host then introduces the guest, Gavin Baker, described as running a firm that invests in private and public markets with "4,000,000,000 under management" (the transcript renders the firm's name as "Atreaties Management"). The opening agenda is the upcoming CoreWeave IPO and NVIDIA, with a mention that Jensen Huang singled out the guest at an NVIDIA event.

The reply narrows the claim: others were also called on, including someone from Sequoia and a sell-side analyst, but "It is technically true. I was the only public equity investor on the buy side that Jensen asked." The speaker says they have known Jensen for twenty-five years and that he is "kind of the same guy he ever was, like, just maybe slightly calmer."

The first substantive question takes up critics' doubts about NVIDIA's balance sheet, in particular its accounts receivable and claims of "round tripping" through investments in customers. The answer rejects the round-tripping story: if NVIDIA had not invested in CoreWeave or other neoclouds, "I don't think it would have impacted their revenues at all at all." In that scenario, the demand would simply have gone to Meta, Tesla, Amazon and Microsoft.

The rationale given is strategic. Three years ago the cloud market was "a very stable three player oligopoly" of Amazon, Google and Microsoft, and having only three big customers is "not great" for NVIDIA. The hyperscalers also each wanted their own custom version of an NVIDIA server. Neoclouds such as CoreWeave bought NVIDIA's standard reference design, which got GPUs into the market faster and was described as "generally smoother and easier to stand them up." Because the big three's share of incremental revenue fell, NVIDIA ended up with "a more fragmented base of buyers who have less power over them."

A participant compares this to Intel's "Intel Inside" program in the eighties and nineties. In this account, Intel probably did not need the program in hindsight, but it supported an ecosystem and spread its buyer base. The guest's argument is restated as: the "quote, unquote, round tripping" is really a way of diversifying NVIDIA's customer base. The answer agrees and adds that the money also sends GPUs to buyers who use NVIDIA's reference architecture. Hyperscalers tweak their servers to resemble their other servers, which "maybe it doesn't run quite as well" as the standard design. The Intel analogy is welcomed as one the speaker "had not thought of."

The branding side of the analogy is then explored. The speaker says Jensen is giving keynotes the way Intel did and branding the names of the units, the way buyers once asked whether a computer had a Pentium. Apple's M4 series is cited as a current case. The conclusion is that some portion of consumers want to know what is inside, and branding "doesn't cost much."

On receivables, the question cites accounts receivable rising from "a billion 5 or so to, like, 5 and a half billion year over year," and a later remark puts cumulative receivables at "like, $23,000,000,000." The answer concedes the point ("It's never good when accounts receivables go up") but argues that this "is the most understandable time for it to happen." The reason given is that the move from Hopper to Blackwell is the biggest product transition in semiconductor history for a business "doing tens of billions of dollars at scale."

The only precedent offered is the iPhone and the "Osborne effect," where buyers hold off in the months before a new model ships. In the speaker's view, NVIDIA kept growing through the transition only because reasoning models "like DeepSeek, which are just so compute hungry" kept demand up.

The physical comparison shows why the transition is so hard. As described:

Hopper rack: about seven feet high, about a thousand pounds, 60 kilowatts ("which is 60 American homes"), air cooled. Blackwell rack: about eight feet tall, four or five feet deep, 3,000 pounds (three times the weight), described as "twice as much power" in "roughly kind of the same footprint," and liquid cooled.

The speaker notes that one of the biggest problems with GPUs is that "they melt," which is why the cooling change matters. The analogy is an iPhone upgrade where you not only change connectors (as in the switch from Lightning to USB-C) but also "put in a new generator, you have to put in a new boiler, and a whole house humidification system." NVIDIA was recognizing Blackwell revenue while the systems were "still kind of getting into customers' hands," which is offered as a large part of why receivables rose faster than sales.

The explanation comes with a stated limit: "If that trend continues past the July, then I would say, hey. That's reason for concern." A participant summarizes that receivables build during a complex transition, peak, and then shrink as they are worked through. When another participant suggests the money is uncollected because the product hasn't been delivered, the answer corrects this: the accounting is complicated, NVIDIA has delivered and recognized the revenue, but customers may be withholding payment "until we get them plugged in and working." The speaker adds, "I mean, who knows what it is?"

The panel lays out the deal: CoreWeave is listing on Friday and the show is recorded on Thursday. It is described as a "Neo Cloud" raising 1.5 billion dollars at a 23 billion dollar valuation, down from a reported 2 billion target. The listed figures are 2 billion dollars in revenue last year ("up 8x"), almost 8 billion dollars in debt, Microsoft at over 60% of revenue ("always a bit of a red flag"), and NVIDIA at 15% of revenue while owning 6% of the company. The questions are whether this is a milestone for an IPO market whose last notable listing was Reddit, whether debt is forcing the listing, and whether the industry "overbuilt" hardware for language models.

The answer is blunt: "So we're definitely not overbuilt yet." The evidence offered is that since DeepSeek R1, China has been "buying every GPU they can"; a large Chinese server maker warned of a coming GPU shortage in China; DRAM prices "go up every day"; and OpenAI said it was limiting its new image-generation service for lack of GPUs.

The speaker predicts Blackwell will be "a very successful product cycle," driven partly by a "prisoner's dilemma": firms spend because they fear a competitor gaining a big advantage, as when Meta might hold back while Google spends and pulls ahead. Another participant says that is "already the case." The speaker agrees and adds that Blackwell is both very different and much better, but new GPUs are hard to get working at first, while Hoppers are already "finely tuned." They compare it to skipping ten years ahead to a new Formula One car: "it takes a while till you drive it as well as the old one."

On CoreWeave itself, the guest acknowledges negative sentiment among investors on X, a reduced price range, and the bear case: "a commodity business, loads of debt, loads of CapEx, you know, one really concentrated customer," and "completely undifferentiated." The counterargument is a retail analogy. A company that runs a thousand well-lit, clean, well-stocked stores across 50 states can build a business worth "well over $10,000,000,000," and very few manage it.

Everybody thinks it's easy, but to synchronize tens of thousands of GPUs where they're melting or cables are being unplugged and, you know, you lose a bunch of training data when it happens, I just think it's not it may not be the commodity that everyone thinks it is