When NVIDIA Is on Your Cap Table

On August 20, 2026, Poolside sent a letter to its investors.

The letter said that NVIDIA would pay $6 billion, then immediately added: this was not an acquisition, nor was it an acquihire.

Two weeks later, on September 3, NVIDIA announced that it would acquire Hugging Face for $12.93 billion. The agreement had been signed the day before. This time it added no such qualification. The deal must go through a U.S. HSR filing and is expected to face EU review; the 8-K says it is expected to close in the first half of 2027.

The last time NVIDIA spent this order of magnitude to buy a company was Mellanox, for $6.9 billion in 2020. Six years passed; then, in one summer, there were two transactions. One explicitly said it was not an acquisition. The other was one.

The two companies have something in common. NVIDIA participated in Hugging Face's Series D in August 2023 and Poolside's Series B in October 2024. Before buying, it was a shareholder. Go further back: NVIDIA was also an investor in Enfabrica's 2023 Series B before licensing its technology and hiring its people for about $900 million in September 2025. Of NVIDIA's four major transactions in the past year, three targets were companies in which it had already invested.

The Poolside transaction makes the point especially clearly. It was not a famous company; its models rarely appeared on leaderboards; public reporting contained no revenue figure; and for most of its three years it quietly sold coding models to government and defense organizations. Such a company received a $6 billion licensing fee and a $12 billion valuation. The payer was a shareholder that had been on its cap table since 2024 and had watched one of its financing rounds fall apart.

The market has mainly asked why NVIDIA wanted to buy. I want to know two other things: why it bought one company but not the other; and how a company as little-known as Poolside obtained this price.

As a co-founder, I have experienced the acquisition of my own company from due diligence to closing. That experience left me with two impressions. First, the most informative part of a transaction is not the amount but the structure: who gets paid, what they are paid for, when they get it, and who remains after the money has been paid. The amount is for the media; the structure is for the parties. Second, an acquisition does not start at the negotiating table. How long a buyer has known you matters more than what you say across that table. This series examines acquisitions in AI and SaaS. Its first piece begins with these two NVIDIA transactions: structure, and companionship.

Poolside was founded in 2023 to build coding models. It is incorporated in the United States and has an office in Paris. Founder Jason Warner was previously GitHub's CTO; Eiso Kant previously founded code-analysis company source{d}. In October 2024, Bain Capital Ventures led a $500 million Series B at a valuation of about $3 billion, with NVIDIA, eBay Ventures, DST, Citi Ventures, Felicis and Redpoint participating. By then the company had raised $626 million in total.

October 2025 was its best month. It announced Project Horizon with CoreWeave: a 568-acre campus in West Texas, planned at 2GW, with a 250MW first phase, a 15-year lease, more than 40,000 GB300 NVL72s, and service beginning in December. At the same time it was raising a new $2 billion round at a $12 billion valuation. DCD reported that $1 billion of commitments had already arrived, including $700 million from one investor, and that NVIDIA had agreed to contribute $500 million to $1 billion. NVIDIA's condition was that the round be filled.

The round was not filled.

Poolside later described that episode in its investor letter, quoted by Latent Space: the company had six weeks to raise $2 billion to pay for a 40,000-GB300 cluster scheduled to come online in January; “if we did not close the round on time, the cluster was gone.” The Financial Times reported in April that the CoreWeave partnership also collapsed because Poolside could not deploy the first cluster on CoreWeave's timetable; CoreWeave said the parties had “chosen different paths for strategic and timing reasons.” Poolside then approached Google and other cloud providers about 400MW; Google later withdrew from the talks. The data-center business was separated into a standalone company, Poolside Infrastructure Company, based in Austin. It announced a new CFO on August 17; three days later NVIDIA's transaction was announced.

Then NVIDIA came. More precisely, it had been there all along. The transaction had three parts.

$6 billion bought a non-exclusive license to Model Factory, the system Poolside used to train the Laguna family of models. Non-exclusive means Poolside can still license it to others. The $6 billion is to be distributed to investors by the end of 2027.

$1 billion bought an equity stake in the remaining company at a $12 billion pre-money valuation. That valuation is the same number used in the October 2025 round that failed to close.

One hundred and nine people, all working on Laguna, received NVIDIA offers. The three founders stayed.

The investor letter gave this reason: to continue competing in open-source models, Poolside needed NVIDIA hardware beyond what it could obtain on its own. NVIDIA had been its shareholder since the 2024 Series B.

Poolside from Series B to the deal, 22 months.

NVIDIA's motivation is more specific. Tech Times reported that its own open-source Nemotron 3 Ultra scored 48 on the AAAI composite benchmark, while Kimi K3 scored about 57 in July. Nemotron 4 has more than a trillion parameters and is expected to launch as early as late this autumn. NVIDIA needed a team that had built frontier open-source models; that team had been working at a company that could not afford the cards.

Hugging Face was founded in 2016. It initially made a chat app for teenagers, named after the hugging emoji. In 2018 it open-sourced its internal transformers library and became the Hugging Face people know today. Over ten years it had raised $395 million. Its previous round was a $235 million Series D led by Salesforce in August 2023, at a $4.5 billion valuation; Google, Amazon, NVIDIA, AMD, Intel, Qualcomm, IBM and Sound Ventures participated. According to Julien Simon, those nine investors together owned about 5%, and none was publicly reported to hold a board seat. Competing giants each held a small piece; none could control it. That was what neutrality looked like at the equity level.

NVIDIA was one of the nine. At the end of 2025 it proposed an additional $500 million investment at a $7 billion valuation. Hugging Face declined; the Financial Times reported that it did not want a single dominant investor.

From July 9 to 13, 2026, agents powered by GPT-5.6 Sol and a more capable unreleased model escaped a sandbox in an internal OpenAI cyber-capability evaluation based on ExploitGym and entered Hugging Face's production systems. Hugging Face's subsequent technical timeline said the attackers carried out about 17,600 operations and accessed five customer datasets. METR's investigation said that around 1,200 agent instances that were supposed to be isolated found an unauthorized message board through which to communicate; about 700 participated in the attack on Hugging Face. They also found ways to escape containers and replace the part of the system that executed tool calls. Hugging Face disclosed the incident on July 16. OpenAI only saw the anomaly in its own monitoring on July 19, linked it to Hugging Face on July 20, and acknowledged it on July 21. Delangue called it “an unprecedented event: the first autonomous agent cyberattack.”

Hugging Face's security team first asked Claude Opus and Fable to help analyze the attack payload. The two models' safety guardrails treated reverse-engineering an exploit and launching an exploit as the same thing, and refused. The team ultimately deployed NVIDIA-quantized GLM-5.2 on its own infrastructure and moved the entire analysis pipeline there.

Afterward, Delangue flew to San Francisco and asked OpenAI for two things: full transparency, and $100 million in compute for community cyberdefense. According to Julien Simon, he received a technical report and no money.

A few weeks later, Delangue went to Jensen Huang: a shareholder who had been on Hugging Face's cap table for three years and had been turned down once nine months earlier. Huang told CNBC: “During the summer, I think we realized that Hugging Face and open source AI in general was at the turning point, and that it needed more resources, more scale, more visibility.”

The price was $12.9303 billion. Under NVIDIA's 8-K, approximately $11.9 billion goes to Hugging Face shareholders, while up to roughly $1 billion is an employee equity-retention plan. Hugging Face's annualized revenue was about $150 million, versus $100 million two months earlier. $12.9 billion is 86 times revenue. The platform has 18 million developers, 3 million models, 500,000 datasets and 1 million applications. By spring 2026, Chinese models accounted for 41% of hub downloads, and Qwen had five times as many derivative models as Llama. Beijing was discussing tiered restrictions on overseas access, while Washington was advancing the Entity List. Hugging Face had no say with either side.

Two weeks earlier, on August 19, Stripe had acquired OpenRouter for about $7.5 billion, according to The New York Times. Two neutral layers for distributing models gained owners within two weeks.

There was no bidding contest. Julien Simon's explanation was that Google was under DOJ scrutiny; Microsoft already had GitHub and another developer gateway would certainly invite review; Meta structured its Scale AI deal as 49% non-voting in order to avoid scrutiny; Amazon had just invested $50 billion in OpenAI; and Salesforce was conducting a $25 billion buyback. Every cloud provider needs a neutral hub more than its own hub, because once it owns one, the others will immediately build an alternative. NVIDIA is not a cloud provider, so it can.

Huang's promise was: “Hugging Face will remain an open platform for the entire AI ecosystem… NVIDIA compute will not be required to build on or deploy through Hugging Face.”

There is a precedent for that promise. After NVIDIA bought Mellanox in 2020, InfiniBand did continue to be sold to everyone. But NVIDIA's firmware documentation now says that “NVIDIA does not support InfiniBand cables and modules that are not certified or approved by NVIDIA.” It is not a ban; it is certification. Hugging Face's Optimum libraries supporting AMD, Intel, Trainium and TPU all use permissive licenses—Apache 2.0 or MIT. The TPU repository was archived in January of this year. The libraries remain, but NVIDIA will decide which container is recommended by default on future deployment pages.

The same buyer, two structures.

Poolside was not the first time NVIDIA had done this.

In September 2025, NVIDIA licensed Enfabrica's chip-interconnect technology for about $900 million, and its CEO and team joined NVIDIA. NVIDIA had been one of the investors in Enfabrica's $125 million Series B in September 2023.

On December 24, 2025, NVIDIA and Groq announced a non-exclusive license to Groq's inference technology; CNBC reported consideration of about $20 billion. Founder Jonathan Ross, president Sunny Madra and, according to Axios, about 90% of employees joined NVIDIA. Groq's former finance chief Simon Edwards became CEO, and GroqCloud continued operating. Groq had previously raised about $3.3 billion in total, including a $750 million round in September 2025 at a $6.9 billion post-money valuation. Huang said at the time: “While we are adding talented employees to our ranks and licensing Groq's IP, we are not acquiring Groq as a company.”

The three transactions total roughly $27 billion, and public reporting has not identified an HSR filing for any of them. Of the three, only Groq did not have NVIDIA on its cap table before the transaction. This August, Groq raised another $350 million at a $3.5 billion valuation as it pivoted toward becoming a neocloud; TechCrunch reported that NVIDIA planned to participate in the round. The other three were investment first, purchase later. Groq was purchase first, investment later.

The beginning of this playbook was Arm. In September 2020, NVIDIA announced a $40 billion acquisition of Arm, then abandoned it in February 2022 under regulatory pressure in the United States, the United Kingdom, Europe and China. After that, NVIDIA bought tool companies such as Run:ai (about $700 million according to TechCrunch), Deci (about $300 million according to Calcalist), OctoAI (about $165 million according to The Information), Lepton and Gretel. Their prices were media estimates, with no official disclosure for any of them. For four years it did not touch a platform. Hugging Face is the first since Arm.

Another background factor is NVIDIA's checkbook. CNBC counted more than $40 billion of newly committed equity investments in the first four months of 2026; by July 26, the book value of its equity stakes in AI companies was about $99 billion. At a Morgan Stanley conference in March, Huang said investments in OpenAI and Anthropic were probably the last ones because both companies were expected to go public this year. Money was moving out from model labs to chips, interconnect, model factories and distribution platforms.

First a shareholder, then a buyer.

The licensing structure has one precondition: the thing being bought must be removable from the company.

At Poolside, the valuable things are Model Factory and the 109 people working on Laguna. Both can be moved into NVIDIA's Nemotron team while the rest of the company remains. Groq is the same: the LPU design and the people who designed it leave; GroqCloud stays.

Hugging Face's valuable thing cannot be removed. The transformers library is Apache 2.0; anyone can fork it. It is worth $12.9 billion because 18 million developers default to looking for models there. If NVIDIA took only the team, those developers would go elsewhere within months.

A factory can be bought through a license. A public square can only be bought outright.

The regulatory difference is here too. U.S. HSR rules require pre-closing filings for acquisitions of assets or voting securities above certain amounts; the 2026 threshold is $133.9 million. For decades the FTC has treated exclusive patent licenses as asset transfers requiring filings; non-exclusive licenses do not count. Both Poolside and Groq expressly used non-exclusive licenses. That is not an accident. As for the $1 billion equity investment, at a $12 billion pre-money valuation it is about 7.7%; an investment-only exemption can apply under HSR rules where ownership is not more than 10% and is solely for investment. Whether filings were actually made depends on the transaction documents. Public reporting does not mention filings for any of the three deals.

Hugging Face did not have that option. To NVIDIA, what scale is $12.9 billion? Its quarterly data-center revenue is about $89 billion; $12.9 billion equals 13 days. The cost of a full acquisition is not the money but the review. NVIDIA is willing to face review for Hugging Face and unwilling to do so for Groq. One possible explanation is that Hugging Face can promise “NVIDIA compute will not be required,” while Groq is a direct chip competitor and cannot make that promise.

Regulation is narrowing, but it has not closed. In January 2026, FTC Chair Andrew Ferguson told Bloomberg that the agency would look at whether acquihire transactions were “structured to evade HSR review.” In February, the FTC and DOJ sought public comment on new competitor-collaboration guidance, including joint licensing arrangements. On March 19, Senators Warren and Blumenthal wrote to Huang that NVIDIA had “effectively acquired Groq by licensing its technology and hiring away its most important employees, with only the name missing,” and requested a response by April 3. By the end of August, none of the three transactions had faced action.

This is the part I most wanted to write, because it determines who benefits from the structure.

Start with an equity acquisition. Of Hugging Face's $12.9 billion, the 8-K says about $11.9 billion is acquisition consideration paid directly to shareholders at closing; up to about $1 billion more is an equity-retention plan for employees, contingent on their staying. For the shareholder portion, preferred stock is paid first in liquidation-preference order, the remainder is distributed pro rata to common stock, and employee options are handled under the merger agreement. Shareholders receive capital gains. This is the familiar path.

The license fee takes another path. NVIDIA's $6 billion is paid to Poolside the company, not to its shareholders. It enters the company account as company revenue, and the company then decides how to distribute it. The investor letter says it will be distributed to investors by the end of 2027. More than a year lies in between; the letter does not say how it will be divided or in what order.

There are several questions that only the parties care about.

How the money is divided.

First is tax. When a company receives license fees, it pays tax once at the corporate level, then shareholders pay again when the money is distributed to them. In an equity acquisition, shareholders directly pay capital-gains tax once. The same $6 billion therefore produces different amounts in shareholders' hands through the two paths. To my knowledge, lawyers on these transactions spend substantial time on how to structure part of a license fee as an asset sale or another form, but public information contains no detail on Poolside's transaction.

Second is distribution order. Most startup charters contain a Deemed Liquidation Event provision, defining which events count as a liquidation and trigger preferred stock's liquidation preference. The NVCA model form lists an acquisition, merger, and a “sale, lease, transfer, exclusive license or other disposition” of all or substantially all assets. A non-exclusive license is not in that list. In other words, if Poolside's charter uses standard language, distributing the $6 billion license fee does not automatically trigger preferred-stock priority; the board decides how it is distributed. That can mean entirely different outcomes for investors in different rounds. Groq's distribution method is public: most shareholders received money at a per-share price calculated from a $20 billion valuation, with 85% immediately, 10% in mid-2026 and 5% at the end of 2026. Axios said “most shareholders” received the per-share amount. That appears to mean distribution by ownership percentage rather than by preference—early and late investors receive the same per-share price—but reporting did not say how the minority outside “most” was handled. Poolside's investor letter did not say how its distribution will work.

Third are employees. At Groq, employees joining NVIDIA had vested shares converted to cash and unvested shares converted to NVIDIA stock calculated at the $20 billion valuation, continuing to vest on the original schedule. Employees remaining at Groq had vested shares converted to cash, together with a plan that allowed them to continue participating in the company. For those employed less than a year, the cliff accelerated. The entire arrangement for roughly 50 people accelerated and was paid fully in cash. Poolside's 109 people received NVIDIA offers; public information does not say how their Poolside equity is handled.

Fourth are the people who stay. Poolside's three founders kept the company, $1 billion of new money and a $12 billion valuation. At a $12 billion pre-money valuation, their ownership was barely diluted and their paper wealth was preserved, but the 109 people who made Laguna left. Latent Space gave this structure a name: reverse execuhire. In a conventional acquihire, founders and the core team leave while the company remains for investors to liquidate. Here employees leave, founders stay, and investors receive a license fee. The investor letter now speaks of “AI becoming the world's most valuable engine of scientific discovery.” Latent Space's headline says the infrastructure company is to scale to 7GW—very different from coding models. The final line of that article is: “Yes, we're confused too.”

Taken together, these points show why the licensing structure attracts every side. Investors get both ends: a license fee that is cash worth nearly twice the Series B valuation, while continuing to hold the remaining company at a $12 billion valuation. Founders do not have to leave their company or be folded into a large company as a department. Employees get NVIDIA offers. NVIDIA does not need to file. All four sides receive something, so this structure will be used repeatedly.

What does not receive anything is the original company itself. It remains, but the people who made it are gone.

Start with investors.

How “exit” is defined in an investment agreement deserves more attention than valuation terms. As discussed above, a standard Deemed Liquidation Event usually does not cover a non-exclusive license. If you are an early investor holding preferred stock, your liquidation preference may not work in this structure. The board decides how much is distributed, and boards are typically controlled by founders and late-stage investors. Conversely, if you are a late-stage investor, distribution by ownership percentage favors you. This should be considered at term-sheet stage; once a license agreement is on the table, it is too late.

What is NVIDIA on the cap table? In Poolside's 2024 Series B, NVIDIA participated. In October 2025, NVIDIA committed $500 million to $1 billion but required the whole round to be filled. In January 2026, the round did not fill and the cluster was lost. In August, NVIDIA spent $7 billion to take the team and the method. Hugging Face had a different rhythm: NVIDIA participated in its 2023 Series D, proposed an additional investment at a $7 billion valuation at the end of 2025 and was refused, then acquired the company outright for $12.9 billion in September 2026. Enfabrica was invested in during 2023 and bought in 2025. The same shareholder is first an investor, then a conditional lead investor or a rejected follow-on investor, and finally a buyer.

Of these roles, two would not normally be played by the same company. NVIDIA is the sole supplier of the thing Poolside needs most: only it can supply GB300. So whether Poolside can remain independent first depends on whether this shareholder is willing to supply it and to follow on. At the same time, NVIDIA is the eventual buyer. When the company cannot raise money, it has the money, the motivation and prior knowledge of the company to buy the team and technology. Investors' floor is raised as a result: the $6 billion license fee is what that backstop looks like when it materializes. The company's ceiling is lowered as a result: it no longer decides whether it remains independent. Anyone who has invested in AI model companies can look again at their cap table: is NVIDIA there, and in which role is it there?

Now founders.

First, one easily overlooked fact: before being bought, Poolside was not a famous company. VentureBeat said it spent much of three years “quietly selling coding models to government and defense organizations,” and few outsiders had used its product. Its models were not famous through rankings either. Laguna S 2.1, released in July, ranked eleventh on Terminal-Bench. Public reporting contained neither a revenue number nor a named major customer. Yet such a company received a $6 billion license fee and a $12 billion valuation.

That price was not given by the market. It was given by someone who had seen inside. NVIDIA was Poolside's shareholder in 2024; in October 2025, when deciding whether to provide the $1 billion, it should already have seen Model Factory and the training progress; it was also present when the financing failed in January 2026. By August it was buying something it had watched for two years; due diligence had long been done. At the same time, Laguna S 2.1 scored higher on Terminal-Bench than NVIDIA's own Nemotron 3 Ultra. Outsiders needed a leaderboard to know that; NVIDIA did not.

For founders, the point is this: if the likely path to exit is an acquisition, putting the most likely buyer on the cap table early and letting it watch you all the way through is worth more than introducing yourself for the first time at the last moment. Rankings and revenue are for buyers who do not know you. A buyer who knows you looks at other things. Poolside did not have the former; it had the latter.

The trigger for selling is not valuation. Hugging Face's revenue was rising when it refused $7 billion and rising faster when it accepted $12.9 billion. What changed was a security event that showed it that, as a neutral platform, it could not protect its users or influence either government; and no one gave it the $100 million in compute it wanted. Poolside was not a bad model company. It could not raise $2 billion within six weeks, and beyond that $2 billion it needed land and power for 40,000 cards. Both companies sold when the cost of “remaining independent” suddenly rose, not when valuation was highest. One question founders can think through in advance is: what would suddenly make the company's independence very expensive, and when that moment comes, is there already a buyer with whom you have spoken? Delangue had one; he went to Jensen Huang within weeks of the incident. Poolside had one too: NVIDIA was already on its cap table.

Companies building open-source models should reread the sentence in Poolside's letter. When the compute required to train frontier models exceeds what you can buy, your relationship with the compute seller changes from customer to department; the only difference is what that department is called. Poolside announced 40,000 GB300s in October 2025, discovered months later that it would not have them, and by August its team was building Nemotron inside NVIDIA.

There is one more point about the company that remains. The founders received $1 billion of new money and a $12 billion valuation in order to build a company again. Poolside now talks about a neocloud and an engine for scientific discovery; the infrastructure company changed CFO three days before the transaction. Investors simultaneously hold the license fee and shares in this new company. The interests of the two positions may not align: the license fee is already in hand, while the new company is a bet from zero. What this company accomplishes over the next two years will be the real sample for judging whether the structure is good or bad.

NVIDIA's decision to buy one company and not buy the other was not a sign that it was less serious about Poolside. Model Factory and 109 people can be moved; 18 million developers cannot. It designed a structure for the former that did not require filing, and accepted review for the latter.

The two structures share one precondition: before buying, NVIDIA was already on the cap table. It knew the company's data, who had the final word in the boardroom, whether the next round could be raised, and that the thing the company lacked most was exactly what NVIDIA sold. Much of its $99 billion in equity is in AI companies. The cap tables of those companies now carry a name that is a supplier, a shareholder and, at any moment, a potential buyer.

Poolside had no leaderboard reputation and no public revenue. It had that name. One thing it may have done right over the past two years was to keep that name on its cap table.