Peter Diamandis opened the segment with a line the show has used before: the United States needs a powerful open-weight model to contend with what is coming out of China. This week, he said on Moonshots EP #283, NVIDIA was "pouring $6 billion into developing an open-source AI model and inference infrastructure" with the startup Poolside, meant to give American developers a domestic alternative to Alibaba, DeepSeek and Kimi. In his framing, a chip company was moving up the stack — from silicon and software to being "a platform provider for openweight ecosystems."
An open-weight model is one whose trained parameters are published for download, so anyone can run it on their own machines and adapt it, rather than renting access through somebody else's service. That distinction carries the rest of the panel's argument, because a model you download has to run on hardware you buy.
Who Poolside is, in Emad's account
Emad Mostaque said he had been talking to investors at a tech conference who had backed Poolside originally. The company was set up by the former chief technology officer of GitHub and others, he said. It started out trying to build a coding model, then moved to an open-source model and "model factory" called Laguna, which he said outperformed Thinking Machines' model when it first appeared.
The trouble, in his telling, was scale. A few months ago Poolside tried to raise $2 billion for a massive Blackwell cluster — a large installation of NVIDIA's current generation of AI chips — and could not. "So they lost that cluster," Mostaque said. "And they were like, this is the table stakes we need." What they did have was "a really great solid open source model for its size."
Poolside had already built a business around running models inside customers' own walls. In a post from May, the company described an enterprise deployment stack in which Dell supplies infrastructure and deployment automation, NVIDIA supplies accelerated computing, networking and AI software, and Poolside supplies the agent platform. The intended buyers are organizations with air-gapped networks, approved-hardware rules or code too sensitive to leave their control; Poolside also argued that long-running coding agents make metered token spending hard to predict, another reason to generate tokens next to your own data. That post describes an existing commercial arrangement with its hardware partners, not the deal the panel was discussing.
Why a chip company would want to give models away
Mostaque described what followed as "this weird Nvidia, um, aqua hire type thing where Nvidia is like, we need to build great open source models" — in order "to increase demand for our technology on the Nemotron stack," NVIDIA's own line of open models and software. He expected more deals: "they're going to be making more, more acquisitions up and down the open source stack to be the leader in open source. Because, again, that drives demand for the GPUs more than anything."
He said the first move had been hiring the team of Ashish Vaswani — one of the authors of the "Attention Is All You Need" paper that introduced the transformer architecture — from his company Essential AI, adding that he did not think it had been announced yet. His prediction for the rest of the field was sweeping: what he called the Nemotron Coalition would absorb the effort, so that "a lot of the classic ones like Mistral and Cohere and others won't be building open source models anymore. They'll be building to the NVIDIA reference design."
Licensing instead of buying
Alex wanted to say "something nice about the American open source community," noting that NVIDIA had already invested about a billion dollars in Poolside before this and was now "turbocharging their own Nemotron community." What struck him was the shape of the transaction, which he called a "hackquisition": the founding team moves across, and the core intellectual property is, in his words, "quote-unquote non-exclusively licensed." His understanding was that NVIDIA is non-exclusively licensing key Poolside IP.
His original read on such deals was that they exist to avoid antitrust scrutiny. He drew a line between two things that look alike: talent acquisitions, which are largely about getting people, and deals structured "at least ostensibly" to avoid antitrust review. In the second case, he said, an acquirer wants to be able to argue: "no, actually, we're just a licensee of Pooleside rather than the acquirer. No, we're leaving a competitive open source model layer."
Dave pushed back on the premise that ordinary acquisitions should be a thing of the past. He said he had taken calls from Mercor and from Orrin — whose Kush Bavaria had been on the podcast a week earlier — both hunting for acquisition targets because the hiring cycle is too slow: "I need groups of three, 10, 15 people that work really well together. I don't care what it costs. Like, send them to me tomorrow." By his inbound, acquisition interest was at an all-time high.
The clock everybody is racing
Dave's explanation for the licensing structures was time, not hostility. The Federal Trade Commission, he said, is "very, very friendly to acquisitions right now," and things move quickly — but on AI timelines "the statutory 30-day review alone is like a lifetime," and a company the size of NVIDIA "is always going to get a second look, which is usually 60, 90 days." So, he said, you assemble whatever deal does not need that review and get on with training the model, then close a conventional deal months later once review is done — his example was Elon Musk closing with Cursor after an HSR review. Liam agreed: "This is just purely juggling the regulatory hurdles and obstacle courses."
The FTC's own description of the process is narrower than a fixed 60- or 90-day clock. It applies only to transactions large enough to require premerger notification under the Hart-Scott-Rodino Act, where both parties generally file and an initial 30-day waiting period runs; certain cash tender offers and bankruptcy transactions use 15 days. The agencies can grant early termination of that wait. If an agency issues a second request for information, the parties cannot close until they have substantially complied and a further waiting period has run — typically another 30 days after both sides comply. How long a review takes therefore depends on whether a second request comes and how fast the companies produce documents, rather than on a single statutory number.
What is left behind
Both Alex and Mostaque were more interested in the remainder than the headline. Alex said he had written off such leftovers as "the carcass of the original company" kept alive only to make the structure defensible, but had started seeing some of them come back to life. He plans to watch "what happens to the part of poolside that did not come to Nvidia," which — not investment advice, he added — "could actually be in some sense even more interesting than the part that goes over to Nvidia."
Mostaque supplied the wrinkle. The remaining company, he said, includes Poolside Infrastructure Company, which is building a 1.2 gigawatt data center. "Which might need GPUs. So they may use some of the money that they get for GPUs. Who knows?"
That pulled the conversation toward verticalization, and Diamandis asked how many layers of the stack each company ends up owning. One panelist offered a limit: the natural pull is at the infrastructure layer rather than the application layer, "for physics and other reasons," since a company offering a frontier model probably also wants to be in the data center, energy, satellite and robotics businesses — the stages closest to the innermost loop. Mostaque split it differently, between innovation and execution, arguing that verticalization suits the execution phase, where you move closer to the silicon and closer to the customer and do not need much better models than you already have.
On the panel's reading, both halves of the Poolside arrangement point the same way: an open model that exists to sell chips, and a leftover company with a gigawatt of planned data center that may go out and buy them.