Nvidia Pays $6 Billion to License the Software That Builds Its AI Models

Nvidia has agreed to pay $6 billion for a non-exclusive license to Poolside's Model Factory — the internal platform the AI coding startup used to build its Laguna family of models — and separately invest $1 billion in the company at a $12 billion pre-money valuation, according to a letter Poolside sent its investors, first reported by Newcomer and independently confirmed by Bloomberg on August 21. The deal brings Nvidia's three-deal total in its non-exclusive license-and-talent arrangement to approximately $27 billion, and for the first time applies that template to the software layer that actually constructs AI models — not just inference hardware or chip interconnects.

The arrangement will send 109 Poolside engineers and researchers to Nvidia via job offers. The three co-founders remain at Poolside. The $6 billion license fee is expected to be distributed to Poolside's existing investors by the end of 2027, converting a startup that raised $626 million over its three-year life into what will function as a substantial venture capital return vehicle. Neither Nvidia nor Poolside has issued a public statement.

The Investor Letter's Revealing Admission

The most consequential detail in the reported investor letter is not the price — it is the explanation for why the deal happened at all. Poolside's founders wrote that at the end of 2025, the company had a six-week window to raise $2 billion to secure a 40,000-chip GB300 cluster that was coming online in January. The raise did not close in time. The company lost the cluster.

The letter describes the constraint plainly: it was "not only capital, it is physical data center space and contracted compute."

This disclosure does more to explain the structural dynamics of frontier AI development than almost any other disclosure from a private lab in recent memory. Poolside had a working model-building pipeline, a proven technical team, an industrial training platform, and institutional investors. What it could not obtain within the hard deadline imposed by data-center provisioning cycles was the compute itself. A $2 billion fundraise at a scale AI startup takes longer than six weeks when the capital is tied to a specific cluster acquisition.

The founders describe three and a half years of being "directionally correct in a race where capital requirements went vertical." They argue that the next generation of frontier models will require not just a 40,000-GPU cluster but "far more than an order of magnitude larger cluster." At that scale, even raising the capital does not solve the problem: physical data center space and contracted GPU capacity become the binding constraint, not the valuation or the conviction of investors.

What the Model Factory Actually Is

Poolside's Model Factory is not a product and is not a model. It is the system Poolside built to manufacture models at scale — closer in architecture to a software CI/CD pipeline or a semiconductor fab automation system than to a training script.

The factory is structured as a versioned, directed acyclic graph of modular components. Those components span every stage of model production: data ingestion, preprocessing, pre-training, supervised fine-tuning, reinforcement learning, architecture ablation, synthetic data generation, quantization, and evaluation. The entire pipeline runs on Dagster for experiment orchestration. Every step is tracked immutably — when a component upstream in the graph changes, everything downstream can be automatically rebuilt from the changed state.

The distributed training codebase inside the factory is called Titan. A PyTorch-based library that Poolside extensively modified from an open-source foundation, Titan handles pretraining at scale across GPU clusters, supports supervised fine-tuning and various reinforcement learning techniques, and serves as the single training entry point for the entire factory. The data ingestion layer, built on Apache Spark and Apache Iceberg, can process approximately 20 trillion tokens per day on baseline compute.

The practical consequence of this design is measurable. When Poolside's co-CEO Eiso Kant described the team's throughput on the Latent Space podcast in July 2026, he reported 10,000 to 20,000 experiments per month with a combined team of fewer than 115 engineers and researchers. Laguna S 2.1 — a 118-billion-parameter, 8-billion-active-parameter mixture-of-experts model trained on 30 trillion tokens — went from the start of pre-training to public launch in under nine weeks. In Kant's framing: "When your factory is built well, the big training run is the easy part."

That speed-to-model capability is what Nvidia is paying $6 billion to license. The Nemotron model family — Nvidia's growing portfolio of open-weight models — is the natural deployment surface for the Model Factory's output. Nvidia is separately working toward a trillion-parameter Nemotron 4, first reported by The Information on August 11. A platform that can run 10,000 ablation experiments per month with a team of dozens is a meaningful accelerant for that roadmap.

What Laguna's Architecture Reveals About Nvidia's Intent

The specific model the Model Factory produced is itself an illustration of the engineering philosophy Nvidia is absorbing. Laguna S 2.1 uses a sparse mixture-of-experts design: 256 routed experts plus one shared expert, with only 10 experts activated per token. With 118 billion total parameters but only 8 billion active on any forward pass, the model's inference cost scales with the small active slice, not with the full parameter count.

That architecture lets a model with the knowledge capacity of a large system operate at the inference cost of a much smaller one. On Poolside's published benchmarks, Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-Bench Multilingual — company-reported figures using the maximum of vendor self-reported, benchmark-author, or third-party leaderboard scores. Independent reproduction of those scores has not been documented, and Poolside's own chart shows Kimi K3 and other frontier models ahead on most rows. What Laguna S 2.1 demonstrates is not frontier-level performance but frontier-competitive efficiency: a 118B mixture-of-experts model small enough to run on a single Nvidia DGX Spark desktop system, holding its own against dense models with far more active parameters.

For Nvidia, the significance is not the benchmark rank. It is the design discipline — the ability to produce compute-efficient models through industrialized experimentation — that the Model Factory encodes.

Nvidia's $27 Billion Pattern and the Template It Is Setting

Nvidia has now deployed the same deal structure three times in under nine months. In September 2025, it licensed chip interconnect technology from Enfabrica for approximately $900 million and brought in key staff, including its CEO. In December 2025, it paid $20 billion to license inference chip design intellectual property from Groq and hired that company's CEO Jonathan Ross and most of its engineering team. Groq remained independent, appointed new leadership, and has since raised $350 million at a $3.5 billion valuation — a round in which Nvidia participated. Now Poolside follows the same pattern, with the founding team staying and the core technical staff moving across.

In each case the license is non-exclusive — meaning the acquired company can theoretically license the same intellectual property to others — and no equity changes hands through merger or acquisition. That structure carries legal significance beyond its elegance: neither the Groq deal nor the Poolside deal triggered the Hart-Scott-Rodino Act's premerger notification requirements, which apply to equity acquisitions above certain thresholds rather than to non-exclusive IP licenses. Nvidia moved at the speed of a commercial contract rather than the timeline of a regulatory review.

Bernstein analyst Stacy Rasgon described the Groq deal structure at the time as potentially keeping "the fiction of competition alive." That observation has aged into a framework. Poolside can technically license the Model Factory to anyone. Whether any other company will pay $6 billion for it, or whether the 109 engineers who built it will remain available to support a competing licensee after joining Nvidia, is a different question.

The Regulatory Overhang

The Poolside deal arrives into a specific regulatory moment. In January 2026, FTC Chair Andrew Ferguson said the agency had begun examining whether reverse acquihire structures were being used to route around merger review. On February 4, Senators Elizabeth Warren, Ron Wyden, and Richard Blumenthal sent a letter to the FTC and DOJ urging investigation of Meta-Scale, Google-Windsurf, and Nvidia-Groq, characterizing them as reverse acquihires that "function as de facto mergers." In March, Senators Warren and Blumenthal wrote directly to Jensen Huang asking whether the Groq deal was structured to evade antitrust scrutiny and set an April 3 deadline for a response.

The UK's Competition and Markets Authority examined Microsoft's structurally similar hiring of Inflection's team and cleared it. No US agency has announced a formal review of any Nvidia deal, and there is no public finding that non-exclusive licensing violates current antitrust law.

The Poolside deal will inevitably be cited in that ongoing debate. Three deals totaling $27 billion, each structured the same way, each resulting in Nvidia gaining strategic capabilities it did not previously control, create a pattern that is harder to characterize as incidental than any single transaction.

What Remains of Poolside

Poolside is not dissolving. The three co-founders retain the independently funded company, which now holds $1 billion in new capital from Nvidia and a $12 billion valuation. The founders wrote in their investor letter that they are "not ready to share the updated vision." The company's infrastructure subsidiary, Poolside Infrastructure Company, is separately building a 2-gigawatt AI campus in Pecos County, Texas — a project called Horizon, announced in October 2025, that predates this deal and has its own recently appointed CEO and CFO.

The Laguna models already released — including Laguna S 2.1, available on Hugging Face under the permissive OpenMDW-1.1 license — remain publicly available. The 109 employees receiving Nvidia offers will decide individually whether to accept. Those who remain at Poolside, or who decline Nvidia's offers, represent the nucleus of whatever Poolside becomes next.

The investor letter offers a philosophical marker for that future. Human-level intelligence in software, the letter argues, will be "fully commoditized by open source models," while superintelligence will not. Poolside's founders divide valuable scientific problems into two kinds: those that are intelligence-bound, which AI will commoditize, and those that are experiment-bound — problems like drug discovery that require real-world feedback loops no amount of intelligence can substitute. They argue the second category is where future value concentrates, and that "AI will become the world's most valuable scientific discovery engine."

How a newly cashed-out AI coding startup with an undisclosed vision and an infrastructure subsidiary becomes a scientific discovery engine is a question that will define the next phase of Poolside's story. For Nvidia, that question is secondary. The chipmaker has already extracted what it came for: the industrialized model-building platform that turns GPU clusters into frontier AI models, and the team that built it.

The compute Poolside could not raise $2 billion fast enough to secure sits inside Nvidia's own infrastructure. The software for using it effectively now does too.