The $6B Deal Nvidia Never Had to File

Introduction

Nvidia was already on Poolside's cap table before it bought the place out. Its venture arm had committed up to $1 billion to the AI coding startup in October 2025, money Bloomberg reported was part of a $2 billion round valuing Poolside at $12 billion pre-money. That round never closed. Poolside lost the 40,000-GPU cluster the money was meant to pay for, the raise collapsed in early April 2026, and Nvidia declined to lead any rescue financing. Four months later it paid $6 billion to license the training system Poolside had built and extended job offers to 109 of its roughly 150 employees, at the valuation it set the previous October.

Nobody at the Federal Trade Commission or the Justice Department had to approve a dollar of it, and nobody did. Both agencies had already flagged this structure on the record. FTC Chair Andrew Ferguson said in January his agency was "beginning to examine these acqui-hires to make sure they are not an attempt to get around" merger review, and the DOJ's acting antitrust chief called that kind of conduct a "red flag" in March. No enforcement action has followed against any of the three deals Nvidia has now closed this way.

The structure moved a step earlier in the process. Nvidia doesn't have to engineer its way around merger review once a deal is on the table. It can sit inside a company's financing, watch that financing fail, decline to catch it, and buy the pieces it wants later at a number it already set. That's run three times in eleven months and roughly $27 billion without a merger notice being filed.

The Cluster Poolside Couldn't Pay For

Poolside's own account of how it got here is blunt. From the shareholder letter, as relayed by Quartz off the Wall Street Journal's review of it: "At the end of last year, we had a 6 week window in which to raise $2 billion dollars to pay for a 40,000 GB300 cluster coming online in January. We didn't close it in time, and we lost the cluster." The company has never published the letter.

Losing the cluster started a chain. The $2 billion round Poolside had been marketing since January collapsed in early April 2026, per the Financial Times and DataCenterDynamics, which also reported CoreWeave walking away from an anchor-tenant deal on Poolside's planned Texas data campus in the same stretch. Bloomberg reported later that month that Nvidia declined to lead a rescue. Poolside had raised $626 million by then, including a $500 million Series B in 2024 at a $3 billion valuation, per TechCrunch.

Then August 20. Newcomer's Eric Newcomer and Tom Dotan broke the terms, paywalled but with the numbers in the free lede: $6 billion for a non-exclusive license to Poolside's "Model Factory" training system, $1 billion of new investment at a $12 billion pre-money valuation, and job offers to 109 people. Bloomberg confirmed the package the next day (via Yahoo Finance). The investor letter, per Quartz, insists it "is not an acquisition and it is not an acquihire." Ten months apart, across a lost cluster and a collapsed round, Nvidia's valuation of Poolside did not move by a dollar.

What $6 Billion Bought Without a Filing

Start with the number that decides whether regulators look at all. The FTC's 2026 threshold notice sets the Hart-Scott-Rodino floor at $133.9 million; the Federal Register notice behind it puts $535.5 million as the level at which a deal is reportable regardless of party size. Enfabrica at over $900 million, Groq at roughly $20 billion, Poolside at roughly $7 billion: every one clears that by an order of magnitude. None required a filing, because HSR covers transfers of assets, voting securities and non-corporate interests, and a non-exclusive license isn't an asset transfer while an offer letter isn't a security.

Model Factory is Poolside's training pipeline. TechTimes describes it training models on whether code actually compiles and passes tests rather than on human preference labels. Co-CEO Eiso Kant said on the Latent Space podcast in July that fewer than 70 people ran "far more than 10,000, maybe 10 to 20,000 experiments a month." Against the roughly 150 headcount Globes reported last December, the 109 offers cover about 73% of the company, though reporting describes them as extended rather than confirmed as accepted.

Six days after committing roughly $7 billion, Nvidia held a two-hour earnings call: $96.22 billion in quarterly revenue, net income more than doubled to $53.95 billion. "Poolside" appears zero times in the transcript. Nemotron came up three times, including when UBS analyst Timothy Arcuri asked Jensen Huang about open models and Huang called them "not a threat, but another source of demand." Zero mentions is the detail I'd have bet against.

Poolside Shipped Laguna S 2.1 in July

A month before the deal broke, on July 21, Poolside released Laguna S 2.1, a 118-billion-parameter open-weight coding model (company release). Its own release claims the model "matches or exceeds models several times its size," naming Nvidia's 550-billion-parameter Nemotron 3 Ultra among the comparators. That's Poolside's own benchmark on its own agent stack, and Tencent's Hy3 and Qwen3.7-Max beat or match Laguna elsewhere in the same table. A month later, Nvidia bought the people and the pipeline that produced it.

Poolside itself is not dead. Kant, co-CEO Jason Warner and co-founder Margarida Garcia all stay outside Nvidia, keeping the name and the fresh $1 billion for what Quartz calls "unspecified research projects." Poolside Infrastructure Company, spun out in January, is building a gigawatt-scale Texas data center; it named a CEO two months before the deal and a CFO the same week. Laguna trained in under four weeks on 4,000 H200 GPUs, so real compute remains, but the team that ran it and the pipeline itself both moved to Nvidia.

Who Benefits

Nvidia gets a proven industrial training pipeline at the exact moment its own open-weight line is behind, without a filing that would drag a company holding roughly 90% of the high-end data center GPU market into formal review. That share figure, the 92% PC GPU number and Nvidia's $5 trillion market capitalization all come from the letter Warren and Blumenthal sent Huang in March. What an HSR filing actually costs a company Nvidia's size is the clock and the disclosure, not the fee.

As for how little the license label constrains anything: at GTC on March 16, three days before that letter went out, Nvidia announced a new inference processor built on Groq's licensed chip technology.

Poolside's investors get what the failed Series C couldn't deliver. The $6 billion license fee is slated for distribution to shareholders before the end of 2027, per Quartz and The Next Web, with no sale to consent to and nobody's position ending. The founders keep the company and the brand, plus an infrastructure subsidiary with its own executives.

Eight Days at the Antitrust Division

On February 4, 2026, Senators Warren, Wyden and Blumenthal wrote to DOJ Assistant Attorney General Gail Slater and FTC Chair Ferguson. Their letter says deals like Meta-Scale AI, Google-Windsurf and Nvidia-Groq "function as de facto mergers... all while apparently attempting to bypass the scrutiny typically applied to mergers and acquisitions," and asks the agencies to block or reverse any that break the law.

Eight days later, Slater was out. Politico reported on February 12 that she was given the choice to resign or be dismissed, after losing the confidence of Attorney General Pam Bondi and Deputy AG Todd Blanche over the HPE-Juniper merger, which has nothing to do with any of this. Omeed Assefi, who had held the job before her confirmation, took it back and gave Reuters the "red flag" line five weeks later. He did not name Nvidia; that framing was Reuters' own, and the story notes Assefi declined to discuss particular companies.

Warren and Blumenthal wrote to Huang on March 19, saying the Groq deal "appears to be structured to evade scrutiny by antitrust regulators," citing the anti-evasion provision at 16 C.F.R. § 801.90 and setting an April 3 deadline for four questions. Nvidia responded through a spokesperson, in the same Bloomberg story that reported the letter: "Nvidia did not acquire Groq, which continues to be a separate and independent business." No public written answer to those questions has surfaced. On Poolside, Nvidia has said nothing publicly at all; The Next Web notes as much, and PYMNTS got no reply to a comment request.

Why the Cap Table Changes the Math

Nvidia has a real alternative explanation here. Its open-model strategy predates all three deals: cheap abundant open models drive compute demand, which is the entire business, and Huang organized a July 2026 open letter against premature restrictions on them. Read that way, Nvidia wanted Model Factory to win a market it's trying to grow, and the license structure came from the seller's side: Poolside's investors wanted liquidity without surrendering the company.

Both readings survive at once. A structure evades HSR review whether or not evasion was the design goal, and only one jurisdiction has actually looked at one: the UK's Competition and Markets Authority examined Microsoft's hiring of Inflection staff and cleared it, as SiliconSnark notes, while no US agency has touched the substance of any Nvidia deal.

What's new with Poolside is the seat Nvidia occupied going in. In Groq and Enfabrica, Nvidia reads as an outside buyer finding a structure that worked. With Poolside it was inside the round whose failure created the distress, chose not to backstop it, then set the price on the far side of the collapse at exactly what it had set before. None of that is illegal, and no court has ruled on whether § 801.90 reaches a non-exclusive license paired with voluntary job offers. It does hand a strategic investor an incentive nobody should want it holding, because the cheapest moment to buy a company's engineers is right after its money runs out.

The Bottom Line

The flat $12 billion is the number to hold onto. Between October 2025 and August 2026, Poolside lost its cluster, its anchor tenant, its funding round and about 73% of its staff to a single buyer, and that buyer's valuation of the company came out identical on both ends of the year, the kind of stability that only happens when one side is setting the price.

The day after the Poolside deal broke, Bloomberg reported Nvidia in early talks with Rebellions, a South Korean inference-chip startup, over arrangements ranging from technology licensing and equity investment up to a full acquisition. That last option is the honest complication here: Nvidia doesn't refuse to buy companies outright; it just hasn't had to in the US, three times running. What would make a fourth one different is a regulator deciding the third one counted.