Independent reporting on artificial intelligence.


The Frontier Wire

Editorials

Open-weight AI just lost its most powerful corporate patron

Jensen Huang made Nvidia open-weight AI's biggest corporate backer and its loudest voice in Washington. The commercial logic behind Nemotron and the Hugging Face deal survives him. The political defence does not, just as export policy turns, and a campaign to reinstate him will not fill that gap.

Open-weight AI has lost the most powerful corporate patron it had. Jensen Huang, who resigned as Nvidia’s chief executive on 11 November 2026, did more than any other executive to put money, models and political weight behind AI systems that anyone can download and run. He agreed to buy Hugging Face for $12.93 billion, built the Nemotron family of open models, organised a coalition of labs to train open frontier models, and in July put his name to a letter warning Washington against restricting open weights. Our view is that the open-model movement is drawing the wrong lesson from his exit. The business case that made Nvidia a patron survives him. What does not survive is the movement’s habit of letting one chip executive make its case in Washington, and the campaign to reinstate him is an attempt to keep that habit rather than replace it.

This is an editorial: the argument is the editors’. The facts it rests on come from Nvidia’s releases, blog posts and SEC filings, the Commerce Department’s rule, congressional statements and named news reports, all listed in the sources. The Frontier Wire did not test any model mentioned here. We take no position on the allegations that preceded Huang’s resignation. They are reported, unproven and under investigation, and our record of his exit sets out what is on the record and what is not.

The argument

  • Nvidia’s open-model spending rests on a commercial logic its own 10-Q spells out: demand for open-source models “promotes use of our products.” A new chief executive inherits that logic.
  • Models already released stay released. Nemotron 3.5 Super VL ships under a licence that grants rights “without restriction.”
  • What the movement loses is political: Huang was its loudest defender in Washington just as export policy turned against compute access for Chinese labs, which built many of the most popular open models.
  • Petitioning for his return ties a public-interest cause to the outcome of a private investigation. The movement should instead get Nvidia’s commitments in writing, separate the open-weights case from the chip-access fight and find more patrons.

What Nvidia built for open weights

The scale of Nvidia’s involvement is easy to underestimate because it arrived in pieces. Put together, it is in our view the largest open-model programme run by any chip company, and by Nvidia’s own count it makes Nvidia the biggest contributor of open models and data to Hugging Face.

CommitmentWhat Nvidia saysDateSource
Hugging Face acquisitionAgreed to buy Hugging Face for $12,930,300,000; “NVIDIA compute will not be required to build on or deploy through Hugging Face”3 September 2026Huang’s post on Nvidia’s blog
Contributions to Hugging Face“More than 500 models” and “more than 250 open datasets”; “the largest contributor of open models and data”3 September 2026Same post
Nemotron CoalitionBlack Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab to build an open base model that will underpin Nemotron 416 March 2026Nvidia release
Nemotron 3.5 Super VLOpen 120-billion-parameter model “with open weights, open datasets and training recipes,” trained on about 26 trillion tokens15 October 2026Nvidia blog
Open-weights letterLetter signed by more than 20 companies, shared by Huang on X, warning against “premature restrictions”24 July 2026CNBC
Compute for open models$29 billion of cloud service agreements that “provide the cloud infrastructure to support our research and development of our open models, such as NVIDIA Nemotron, Cosmos, and GR00T,” and autonomous vehicle softwareAs of 26 July 2026Nvidia 10-Q

The last row matters most for anyone worried about money. The $29 billion figure covers autonomous vehicle software as well as open models, and Nvidia does not split it. Even so, it shows that open models are not a side project funded from a marketing budget. They sit inside multi-year compute commitments that a chief executive cannot unwind quickly, whoever holds the job.

Why the patronage was personal

The money belonged to Nvidia. The advocacy belonged to Huang. He framed open models as a matter of national strength rather than corporate strategy, and he did so in places where policy is made.

At the Center for Strategic and International Studies on 3 December 2025, he argued that “without open source, startups can’t thrive, university researchers can’t do research, you can’t teach AI, scientists can’t use AI.” When Nvidia launched the Nemotron Coalition in March, he called open models “the lifeblood of innovation and the engine of global participation in the AI revolution.” In the Hugging Face post he wrote that open weights “help ensure that AI leadership is distributed across companies, institutions and communities.”

The July letter showed what his name could do. CNBC reported that it was released by Nvidia, Microsoft, Meta, Palantir and more than 20 other companies, and that it argued: “Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect.” Semafor reported that OpenAI later backed the letter and that Anthropic did not sign it. Few other executives could have assembled that list.

People around the movement describe his role the same way. MIT Technology Review wrote on 16 November that Huang “had become a sort of ringleader for those who defend open source to policymakers against concerns that such models come with no guardrails and cede AI leadership to China.” Su Lian Jye of Omdia told the South China Morning Post that “Huang has been the singular driving force behind Nvidia’s aggressive, high-conviction bets.”

What the exit does not change

Start with the reason Nvidia funded open models in the first place. Its 10-Q for the quarter ended 26 July 2026 puts it plainly: “We believe the demand for open-source foundation models and applications promotes use of our products worldwide. If these models and applications are primarily deployed on competitors’ platforms, they could reduce demand for our offerings.”

That sentence explains Nvidia’s interest better than any keynote. Open models let thousands of companies run AI on their own hardware rather than renting it from a few closed-model labs, several of which are designing their own chips. The South China Morning Post quoted Rui Ma of Tech Buzz China making the same point: Nvidia benefits when the industry “isn’t dominated by a handful of giant companies that can afford to develop their own chips.” That logic does not depend on who is chief executive. It depends on Nvidia selling GPUs, which it will keep doing under Colette Kress or anyone else.

The first evidence after the resignation points the same way. On 16 November, four days after Nvidia announced Huang’s departure, the company released Nemotron for Science, which pairs the open Nemotron 3.5 Super VL model with LitmusBench v0.2, an open dataset of nearly 50,000 chemistry questions, and post-training recipes. The release quotes Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, not Huang. The work was presumably planned well before his exit. Its timing still shows that the open-model pipeline did not stop when he left.

The models already released cannot be taken back. The Hugging Face card for Nemotron 3.5 Super VL lists the OpenMDW-1.1 licence. Its text grants permission “to deal in the Model Materials without restriction, including under all copyright, patent, database, and trade secret rights.” It says nothing about revocation by the licensor. The only termination clause applies to a licensee who sues claiming the materials infringe a patent or copyright. A future Nvidia leadership could stop publishing new models. It could not withdraw these. Our survey of open-weight licensing found that most flagship releases use similarly permissive terms such as Apache 2.0 or MIT, while the exceptions carry user caps, revenue gates or remote kill switches. Some of Nvidia’s own earlier Nemotron licences ended a user’s rights if a safety guardrail was bypassed without a substitute.

The Hugging Face commitments are also written down, at least in Nvidia’s own announcement. Huang’s post says Hugging Face “will remain an open platform for the entire AI ecosystem,” will keep supporting “open source and open weight models from across the ecosystem, from every model builder,” and will continue to support “multi-cloud and multi-accelerator development and deployment.” TechCrunch reported that the deal is still scheduled to close in the first half of 2027, and that, according to Reuters, the Justice Department has been asking how it would affect competition.

What the exit does change

Nvidia’s interest survives. Its willingness to spend that interest in public may not.

Huang argued for open weights while also arguing that cutting China off from American technology would hurt the United States. The South China Morning Post noted that he “has repeatedly argued against cutting China off from American AI technology.” Those two positions were linked in practice. MIT Technology Review describes China as the place “where the most powerful and popular open-source models are created,” and reported that renting Nvidia chips remotely in countries such as Malaysia was “a workaround that has helped China’s open-source models get so good.”

That link is now a liability. On 16 November, the Commerce Department’s Bureau of Industry and Security released a rule that it says was “written by and at the direction of the Department of Treasury and the National Security Council.” It requires licences for chips going to data centres anywhere that provide more than 12 million aggregate TPP to companies headquartered in China or other Country Group D:5 destinations, and adds licence requirements for five Southeast Asian and Latin American countries. Our guide to the remote-access rule covers the details. MIT Technology Review argued that the rule “could mean the models popular on Hugging Face will soon stop getting smarter.”

The political mood has moved with it. Three House committee chairs said on 12 November that the case “is a prime example of why Congress must immediately pass the AI OVERWATCH Act, Chip Security Act, MATCH Act, and Remote Access Security Act.” Rui Ma told the South China Morning Post: “The political balance has now shifted considerably towards those advocating tighter restrictions. It could mean even greater scrutiny of Chinese open-weight models and restrictions on their availability and use.”

A report adds to the pressure without settling anything. MIT Technology Review and the South China Morning Post both reported that Will DePue, a former OpenAI researcher, published findings after Huang resigned claiming that leading Chinese labs trained flagship models on restricted Nvidia Blackwell chips. MIT Technology Review said an independent analysis had not confirmed the results and that the findings do not establish that Nvidia helped anyone obtain chips. Unverified or not, claims like this are what an open-weights advocate in Washington now has to answer.

A chief executive who is also the company’s finance chief, running a company under congressional and Treasury scrutiny, has every reason not to argue these points in public. Kress may continue every open-model programme Huang started and still decline to defend them before Congress. That is the loss.

The reinstatement campaign is the wrong fight

The open-model world has responded by trying to get Huang back. MIT Technology Review reported that a petition to reinstate him lists, among its signatories, “higher-ups from Hugging Face” and people from Andreessen Horowitz and Sequoia Capital, “a particular slice of the tech world” with “a vested interest in the development of open-source AI.” On 17 November, Huang said that “if the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.” CNBC had reported the day before, citing a person familiar with the matter, that the board is not considering it.

Our view is that the movement should stop. There are three reasons, and none depends on any judgement about Huang’s conduct.

It ties a public cause to a private process. Whether Huang returns depends on an investigation by Nvidia’s independent directors, on a transcribed interview before three House committees scheduled for 1 December, and on federal inquiries into Megaspeed. Open-model advocates have no part in any of those. Linking the future of open weights to their outcome hands control of the cause to people who are not weighing it.

It confirms the critics’ description. One charge against open weights is that they serve a narrow commercial coalition. MIT Technology Review’s account of the petition makes that point for them: it describes a significant faction of the signatories as platform executives and venture investors, and notes that many open-source advocates stand to profit. A movement that claims to serve “startups, businesses, universities and public institutions,” in Huang’s own phrase, should not look like a shareholder lobby.

It concedes that the case needs him. If open weights are good policy only when Jensen Huang argues for them, they were never as strong a case as their supporters claim. The arguments in the July letter, that closed models “can be breached, misused, or fail in ways that outsiders cannot detect” and that concentration “compounds that risk,” do not lose force because the person who shared them has resigned.

Clément Delangue, Hugging Face’s chief executive, struck a better note on the day Nvidia announced Huang’s resignation. “This whole thing sucks, and I wish none of it had happened. It’s a really sad day! But I want to make one thing clear: we’ll keep fighting for open-source AI, harder than ever,” he wrote on X, as TechCrunch reported. Hugging Face is due to become a subsidiary of Nvidia if the deal closes. That gives Delangue more reason, not less, to make the case under his own name.

If open weights are good policy only when Jensen Huang argues for them, they were never as strong a case as their supporters claim.

The Hugging Face question

The acquisition is where Huang’s departure matters most, because it concentrates the open-model world’s infrastructure in one company at the moment that company has lost the executive who made the promises about it.

By Nvidia’s account, more than 18 million developers, researchers and creators use Hugging Face to share more than 3 million models, 500,000 datasets and 1 million applications, and more than 200,000 companies use it to discover and deploy AI. Those are Nvidia’s figures, published in the acquisition post. TechCrunch put the trade-off plainly: Nvidia pledged that the platform would preserve open support, but the deal “would also allow the company to control a major distribution channel for AI models.”

Huang’s answer to that concern was personal and specific. He wrote that “developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want and the computing platforms they want.” He described Delangue approaching him “as he considered the next chapter of Hugging Face.” Those assurances were credible largely because of who gave them: a founder who had spent years arguing that open models were good for Nvidia because they kept AI from being concentrated in a few hands.

None of that binds his successor. A blog post is not a contract with users, and the acquisition has not closed. A future Nvidia leadership under investor pressure could keep every promise in the post and still change how Hugging Face ranks, promotes or optimises models, in ways that favour models tuned for Nvidia hardware. We are not predicting that. We are pointing out that the safeguard most open-model users rely on today is the word of a chief executive who no longer holds the job.

The patrons already in the room

The open-model camp does not have to start from nothing. Nvidia’s own programmes have already drawn other companies in, and some of them have said why in public.

OrganisationRole in Nvidia’s open-model workStated reason, in its own words
Mistral AICo-developing the coalition’s first base model with Nvidia“Open frontier models are how AI becomes a true platform.” (Arthur Mensch)
PerplexityCoalition member; customising Nemotron 3.5 Super VL for search and agents“Open models make AI more accessible at scale.” (Aravind Srinivas)
Thinking Machines LabCoalition member contributing data collaboration“We believe in AI that is adaptable, collaborative and broadly accessible.” (Mira Murati)
SarvamCoalition member contributing sovereign language work“Open models make this possible by giving builders the freedom to adapt frontier capabilities to real-world needs.” (Pratyush Kumar)
CursorCoalition member contributing evaluation datasets“When frontier models are accessible and transparent, developers everywhere can help shape how this technology evolves.” (Cursor’s co-founder and chief executive)

The quotes come from Nvidia’s March coalition release. Nvidia’s October Nemotron post adds a longer list of organisations it says are customising or integrating Nemotron 3.5 Super VL, including Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys, with Scale AI post-training it for clients’ agents and AI21 for cybersecurity. Those are Nvidia’s descriptions of its customers, not independent accounts.

The table makes a simple point. The companies that benefit from open frontier models include labs, search engines, developer tools and enterprise software vendors, not only one chipmaker. Several of them already contribute work to a shared open model. The question after Huang is whether any of them will contribute money and public argument on a similar scale, or whether they will continue to let Nvidia carry both.

What the open-model camp should ask for instead

Patrons come and go. Commitments in writing, a cleaner argument and a wider base outlast them. We propose four steps. They are the editors’ proposals, not commitments any company has made.

Get Nvidia’s commitments on the record from its current leadership. The most important open-model promises Nvidia has made appear in a blog post signed by a chief executive who no longer works there. Nvidia’s board and Kress should restate them in their own names: that Hugging Face will stay open to every model builder and every accelerator, and that Nvidia compute will not be required. If the Justice Department’s review of the acquisition ends in conditions, those promises are the obvious candidates to make binding. Open-model advocates should say so to the department now.

Separate the case for open weights from the fight over chip access. The July letter defended open weights as a matter of competition, transparency and resilience. It did not need to defend Chinese labs’ ability to rent restricted chips through third countries, and the movement should not let the two be bundled. The new rule restricts compute for companies headquartered in China. It does not restrict publishing or downloading open models, and it removed the AI Diffusion Rule’s control on model weights. Advocates who accept the export-control premise and argue for open weights on their merits will be heard by a Congress that would ignore them if they appear to be defending a loophole.

Widen the base of patrons. Nvidia was the largest single sponsor, but it was never the only one with a reason to sponsor. The Nemotron Coalition’s eight members already contribute data, evaluations and expertise to a shared open model. Microsoft, Meta and the other companies that signed the July letter said that America’s AI leadership “will be judged not by one frontier AI model, but by whether the United States builds a strong, open ecosystem.” They can demonstrate that with funding and releases, not only signatures. A movement with five corporate patrons is far harder to weaken than one with a single dominant one.

Publish the evidence. The open-weights argument is often made in the abstract. It would be stronger with numbers that the public can check: how many organisations post-train open models, what they use them for and what safety incidents have been traced to them. Nvidia’s own releases describe customers post-training Nemotron for credit analysis, cybersecurity and scientific work. Those are vendor claims. Independent tallies would persuade more than any letter.

The counterargument

The strongest case against our position comes from those who think open weights should lose ground. The Commerce rule states the concern: access to advanced computing “can be used to train and run advanced SI models capable of supporting military- and intelligence-related activities,” including “more effective weapons of mass destruction” and “coordinated cyber-espionage.” Critics argue that every capable open model is also available to the governments the export controls target, and that an industry champion who opposed those controls was a cost to national security, not a benefit to it.

We accept the first half of that concern and reject the conclusion. Export controls on compute and openness of weights are different instruments with different targets. The new rule is aimed at computing capacity for training and running models, not at publication, and it removed the earlier rule’s control on model weights in favour of destination- and end-use-based controls. A policy that limits where advanced chips go while American companies keep publishing open models is internally consistent. Its weakness was that open-model advocacy had come to depend on someone who opposed the first half. That dependence is what the movement should end.

There is also a commercial objection: that we are too sanguine about Nvidia’s spending. Rui Ma told the South China Morning Post that Nvidia will “probably prioritise their core business over other more tangential objectives like developing open-source models or funding them.” She may be right. Nvidia’s interim leadership has not said anything about the programme either way. Its third-quarter results call, scheduled for 17 November, is the first chance for it to do so; no results release had appeared on Nvidia’s newsroom or in its SEC filings when this editorial was written. But the more Nvidia’s support depends on its own balance sheet rather than on its founder’s convictions, the more important our second and third steps become.

What to watch

  • The first earnings call under Kress. Whether Nvidia’s interim chief executive mentions Nemotron, Hugging Face or open models at all, and whether she frames them as strategy or as cost.
  • Nemotron 4. Nvidia said in March that the coalition’s first model would underpin the Nemotron 4 family. Its release would be the clearest signal that the programme continues at full scale.
  • The Hugging Face review. Whether the Justice Department clears the deal, conditions it, or challenges it before the planned close in the first half of 2027.
  • The 1 December interview. What Huang tells House investigators about Nvidia’s compliance, and whether Congress moves the bills its committee chairs named.
  • New patrons. Whether any of the July letter’s signatories announce open-model funding on a comparable scale.

Open-weight AI was never Jensen Huang’s personal project, even if he was its most visible champion. The best tribute its supporters can pay to what he built is to show that it does not need him.

Sources

  1. NVIDIA: NVIDIA Announces CEO Transition (12 November 2026)
  2. NVIDIA blog: NVIDIA to Acquire Hugging Face, by Jensen Huang (3 September 2026)
  3. NVIDIA blog: NVIDIA Expands Nemotron Family With Super VL and Bolt Models (15 October 2026)
  4. NVIDIA: NVIDIA Launches Nemotron Coalition of Leading Global AI Labs (16 March 2026)
  5. NVIDIA: NVIDIA Introduces Nemotron for Science (16 November 2026)
  6. Hugging Face: NVIDIA Nemotron 3.5 Super VL 120B-A12B model card
  7. OpenMDW License Agreement, version 1.1
  8. NVIDIA Form 10-Q for the quarter ended 26 July 2026
  9. CSIS: Nvidia's Jensen Huang on Securing American Leadership on AI (3 December 2025)
  10. BIS: BIS Issues Rule Addressing Remote Access to American Super Intelligence Technology (16 November 2026)
  11. House Select Committee on China: Chairmen send second request to Jensen Huang and NVIDIA (12 November 2026)
  12. MIT Technology Review: The open-source movement wants Jensen Huang back (16 November 2026)
  13. TechCrunch: Hugging Face CEO vows to keep fighting for open source AI after Jensen Huang's exit (12 November 2026)
  14. CNBC: Nvidia, Microsoft, Meta warn against premature restrictions of open-weight models (24 July 2026)
  15. Semafor: Tech leaders back open-source AI (26 July 2026)
  16. South China Morning Post: How Jensen Huang's downfall could reshape the AI race (13 November 2026)
  17. CNBC: Former Nvidia CEO Jensen Huang says he's open to returning in first statement since resignation (17 November 2026)
  18. CNBC: Nvidia board is not considering bringing Jensen Huang back as CEO, source tells CNBC (16 November 2026)

Questions readers ask

Is this article news or opinion?

Opinion. It is an editorial by The Frontier Wire's editors. Facts about Nvidia's open models, its filings, the Hugging Face acquisition and the Commerce rule come from the primary documents and named news reports listed in the sources. The recommendations are the editors' own.

Will Nvidia stop releasing open models now that Huang has gone?

Nothing Nvidia has published says so. On 16 November 2026, four days after it announced Huang's resignation, it released Nemotron for Science, built on its open Nemotron 3.5 Super VL model, with a new open chemistry dataset. Its latest 10-Q says demand for open-source models promotes use of its products. Analysts quoted by the South China Morning Post say the spending could face closer investor scrutiny.

Can a new Nvidia CEO take back models already released?

Not under the licence on Nemotron 3.5 Super VL. Its Hugging Face card lists OpenMDW-1.1, which grants permission to deal in the model materials without restriction. The only termination condition in the text applies to a licensee who sues claiming the materials infringe a patent or copyright.

Does The Frontier Wire take a position on the allegations against Huang?

No. The intelligence claims reported in the press are unproven, disputed by China and under investigation by Nvidia's board and US authorities. This editorial argues only about what his departure means for open-weight AI.

More editorials