Not an Acquisition: Nvidia Pays $6 Billion for Poolside's Model Factory and 109 Engineers - August 22, 2026

Not an Acquisition: Nvidia Pays $6 Billion for Poolside's Model Factory and 109 Engineers Nvidia licensed Poolside's Model Factory for $6 billion, invested another $1 billion at a $12 billion pre-money valuation, and made job offers to 109 of the fewer than 115 people who built the company's Laguna models. Poolside's investor letter insists this is neither an acquisition nor an acquihire, and it is the third deal Nvidia has structured this way after Groq and Enfabrica. Chris and Laura dig into what a non-exclusive license really transfers, why a lost 40,000-chip cluster forced the deal, and what it means when the company that controls the compute becomes the most likely buyer. Hosted by Chris and Laura. The DX Today Podcast brings you daily deep dives into the most consequential stories in the AI ecosystem. #ArtificialIntelligence #Nvidia #AIChips #TechMergers #OpenWeightModels

Transcript

SPEAKER_01 0:00

Welcome to the DX Today Podcast, your daily deep dive into the AI ecosystem. I'm Chris, and joining me as always is Laura. Today we are digging into a deal that landed this week, which, depending entirely on how you read it, is either the most creative piece of corporate structuring in this industry right now, or an acquisition wearing a very thin costume.

SPEAKER_00 0:19

I have been waiting all week to get into this one because the dollar figures are enormous. The legal language wrapped around them is doing an almost heroic amount of work, and the pattern behind it is bigger than this single transaction.

SPEAKER_01 0:31

Let's start with the raw shape of it, because people need the numbers in front of them before any of the interpretation makes sense. Walk us through what actually changed hands here.

SPEAKER_00 0:41

Nvidia is paying $6 billion to license something Poolside calls its model factory, which is the internal software system that company built to actually produce AI models from scratch. Separately, NVIDIA is investing $1 billion at a $12 billion pre-money valuation. So roughly $13 billion once the money lands.

SPEAKER_01 1:01

So far, that sounds like a large but fairly ordinary technology licensing arrangement bundled with a growth equity check. Two companies doing business at scale. What is the part that makes people raise an eyebrow?

SPEAKER_00 1:12

The third component, NVIDIA is extending job offers to 109 poolside employees, specifically the people who worked on the Laguna model family. And that number is where this whole story quietly turns into something else.

SPEAKER_01 1:25

Why does 109 matter so much? On its own, that sounds like a sizable hiring wave, but plenty of large companies hire 100 people in a quarter without anyone writing a word about it.

SPEAKER_00 1:36

Because fewer than 115 people in total, across all of engineering and research, built those models. So 109 offers is not a hiring wave at all. That is essentially the entire team walking out the door together.

SPEAKER_01 1:49

That reframes it completely. Nvidia is buying the factory, the blueprint for running the factory, and then hiring almost every single person who knows how the factory actually works in daily practice.

SPEAKER_00 2:01

And here is the sentence that makes this newsworthy rather than routine. The letter pool side sent to its own investors says, in plain language, this is not an acquisition and it is not an aqua hire.

SPEAKER_01 2:13

I want to sit with that for a second, because that is a remarkably confident thing to put in writing when you have just described transferring the technology, the method, and nearly all of the people who built both.

SPEAKER_00 2:24

There is a defensible technical argument underneath it. To be fair, the license NVIDIA bot is non-exclusive, which means poolside keeps the right to use that same system itself and to license it to other parties. An acquisition takes the asset away entirely, and this does not.

SPEAKER_01 2:40

That is a genuinely meaningful distinction, and I don't want to wave it off. A non-exclusive license means two parties can walk away from the table holding the same capability at exactly the same time.

SPEAKER_00 2:51

Right, and the company continues to exist as an independent entity. All three co-founders are staying. A new chief executive was appointed roughly two months ago, and a new chief financial officer arrived just this week. So somebody is clearly planning a next chapter.

SPEAKER_01 3:07

Okay, but let me push on that, because a company can technically continue to exist while having been functionally hollowed out. What is actually left inside poolside once those 109 people are gone?

SPEAKER_00 3:18

What remains is a balance sheet, a leadership team, a very large infrastructure project, and the legal right to use technology that almost nobody left in the building has hands-on experience operating at scale.

SPEAKER_01 3:30

The infrastructure piece is worth pausing on because I suspect a lot of listeners have not tracked how physical this company became over the last year. Tell us about the data center side of the business.

SPEAKER_00 3:40

Poolside spun out a separate infrastructure company back in January, and that entity is still building a 1.2 gigawatt data center campus in Texas. For scale, that is genuinely utility-sized, roughly the electrical demand of a mid-sized city. And the original version announced with Core Weave was pitched at two full gigawatts.

SPEAKER_01 4:00

Now I want to get to the part of the story I find most revealing, which is the reason Poolside itself gave for doing the deal in the first place. What went wrong?

SPEAKER_00 4:09

Compute. Specifically, Poolside lost access to a planned 40,000 chip cluster of NVIDIA GB300s after failing to raise $2 billion inside a six-week fundraising window. That is a brutal deadline in any environment.

SPEAKER_01 4:24

Six weeks to raise $2 billion is punishing, even in this funding climate. And missing it apparently cost them the exact hardware allocation that the entire company strategy was built around.

SPEAKER_00 4:34

And their own analysis of next year was even starker. Leadership calculated that staying at the frontier would require a cluster far more than an order of magnitude larger than the one they had just lost. They named three constraints in that letter, and the specificity matters. Capital, contracted compute, and physical data center space.

SPEAKER_01 4:53

Here's what strikes me as uncomfortable about the whole sequence. The scarcity that squeezed poolside was scarcity of NVIDIA hardware. And the company that ultimately resolved the squeeze by writing the check was NVIDIA.

SPEAKER_00 5:05

You have put your finger on exactly the dynamic that has people talking this week. The supplier controlled the constraint. The constraint became existential, and then the supplier arrived as the buyer of the outcome.

SPEAKER_01 5:16

I want to be careful not to imply anything deliberate there, because nobody has shown that.

SPEAKER_00 5:28

And it is not a one-time event, which is what turns this from an anecdote into a pattern worth studying. This is the third deal Nvidia has done with essentially this exact shape in a matter of months.

SPEAKER_01 5:39

Lay out the pattern for us, because seeing the repetition is what makes the strategy legible rather than looking like a series of unrelated opportunistic transactions that happen to share a buyer.

SPEAKER_00 5:49

In May, NVIDIA did roughly $20 billion with Grok for inference technology and senior engineers. There was a similar arrangement within Fabrica at around $900 million, plus comparable deals involving Kumo AI and SCEDMD.

SPEAKER_01 6:04

$20 billion for Grok is a staggering figure to move through a structure that most observers would not immediately recognize as a merger. Did Grok survive that in any meaningful sense afterward?

SPEAKER_00 6:16

It did, and this is the counterargument people should take seriously before writing an obituary. After the Nvidia deal, Grok raised $650 million at a $3.5 billion valuation. So the template has at least one case where the other side kept operating and kept raising afterward.

SPEAKER_01 6:33

Give me those numbers because I suspect the people who wrote the original checks into this company have a very different emotional read on this outcome than the commentators watching from outside do.

SPEAKER_00 6:44

Poolside raised roughly $626 million in total across its entire history, including a $500 million round in the fall of 2024. The distribution going back to those investors is $6 billion, approaching a tenfold return. And the company still keeps the non-exclusive license and a billion dollars of fresh capital.

SPEAKER_01 7:05

So the devil's advocate position is that everybody wins here. Investors get paid handsomely, the engineers get NVIDIA offers, the company gets recapitalized, and NVIDIA gets the capability it wanted without a formal merger review.

SPEAKER_00 7:19

That last clause is where the entire debate actually lives, though. Avoiding a merger review is not a neutral side effect of the structure. For a great many observers, it is the whole point of the structure.

SPEAKER_01 7:31

Let's talk about what NVIDIA is actually getting because I don't want us to treat this as pure financial engineering. The Laguna models themselves are a real technical asset. How good are they in practice?

SPEAKER_00 7:40

Genuinely good. The most recent release is a 118 billion parameter mixture of experts model with roughly 8 billion active parameters per token, and it runs on a single desktop class NVIDIA machine. It scores around 70% on Terminal Bench and just under 60% on SWE Bench Pro.

SPEAKER_01 7:58

How does that stack up against the closed frontier systems, though? I assume there's still a meaningful gap between the best open weight coding model and what the leading commercial labs are shipping today.

SPEAKER_00 8:09

There is a gap. The closed offerings still lead by roughly 10 to 15 percentage points, so nobody is claiming parity here. But the training story is what makes this genuinely interesting for a chip company.

SPEAKER_01 8:21

Tell me about the training story, because if the headline is that you can get near frontier coding performance without a hyperscale budget, that fundamentally changes who gets to participate in this market.

SPEAKER_00 8:31

They trained it in under four weeks on 4,000 H200 graphics processors and released the weights publicly under an open license managed through the Linux Foundation so anyone can download and run it.

SPEAKER_01 8:43

Four weeks on 4,000 chips is a fundamentally different order of investment than what we usually discuss on this show. And it explains why the recipe itself was worth $6 billion to a buyer.

SPEAKER_00 8:54

And that is precisely what NVIDIA bought. Not the finished model, which is already public and freely downloadable by anyone, but the repeatable industrial process for producing models like it over and over again.

SPEAKER_01 9:06

There is also geopolitical framing that Poolside leaned into pretty hard. Positioning these models as a Western counterweight to the Chinese open weight labs. Does that framing actually hold up under scrutiny?

SPEAKER_00 9:18

It holds up better than I expected. They pointed out that no Western lab had shipped an open weight model in that parameter class for 11 straight months while the Chinese labs kept releasing on a steady cadence.

SPEAKER_01 9:29

So one plausible reading of this entire transaction is that NVIDIA just took ownership of the most credible Western open weight model pipeline at a moment when almost nobody else was seriously building one.

SPEAKER_00 9:40

And NVIDIA already runs its own open model line and has talked publicly about pushing toward a trillion parameter open model. So this slots directly into a strategy that was already very much underway.

SPEAKER_01 9:51

Which raises the question I keep circling back to. Nvidia sells the shovels to everyone in this gold rush. What happens when the shovel company starts showing up at the dig site with its own crew?

SPEAKER_00 10:03

That is the tension nobody has resolved yet. Every lab buying NVIDIA hardware now buys it from a company that also ships free models, competing directly with the products those labs sell, and that sees the demand signals and allocation cues of the entire market simply by being the supplier to all of it.

SPEAKER_01 10:19

Spell that out because that is the practical takeaway for anyone building in this space, rather than just watching the scoreboard from a comfortable distance and enjoying the drama.

SPEAKER_00 10:28

If your survival depends on securing a compute allocation you cannot fund on your own, then your most likely buyer is not another lab and not a hyperscaler. It is the company that controls the allocation.

SPEAKER_01 10:40

That is a genuinely uncomfortable conclusion. And it suggests the competitive dynamics in this entire market are increasingly being set at the hardware layer rather than at the model layer or the product layer.

SPEAKER_00 10:52

And notice how quiet the regulatory response has been so far. Three deals tens of billions of dollars, hundreds of engineers moved between companies, and very little formal scrutiny of the pattern as a pattern.

SPEAKER_01 11:04

So, what should people actually watch over the next few months to know whether this template holds, or whether it starts drawing attention from people who carry subpoena power?

SPEAKER_00 11:12

Three things. Whether Poolside ships anything meaningful without that team, whether regulators begin treating repeated licensing deals as a reviewable pattern rather than isolated events, and whether a fourth company takes the same deal.

SPEAKER_01 11:25

I would add a fourth marker to that list. Watch whether other chip suppliers copy this playbook, because if this becomes the standard way hardware companies absorb software talent, the industry map redraws itself very quickly.

SPEAKER_00 11:37

Agreed completely. The thing I keep coming back to is that six billion dollars did not buy a product or a customer base. It bought a method and methods compound in a way that finished products simply do not.

SPEAKER_01 11:49

That is a sharp place to land it. A finished model has a shelf life measured in months, but the ability to keep producing them is the asset that actually holds its value over time.

SPEAKER_00 12:00

And whoever owns that capability, plus the hardware that everyone else needs in order to exercise it, is holding a position in this market that is going to be extremely difficult for anyone to challenge.

SPEAKER_01 12:11

That's all for today's episode of the DX Today podcast. Thanks for listening, and we'll see you next time.