Simon JarnheimerNVIDIA / ESSAY
Chapters
Cobalt architectural illustration of a technology campus with Nvidia signage

Essay

Nvidia and the
price of
staying open

An essay on Nvidia’s proposed acquisition of Hugging Face, and what it could mean for the independence of open source.

Read the essay

Nvidia signed a definitive agreement on 2 September 2026 to acquire Hugging Face. The next day Jensen Huang, Nvidia’s chief executive, named the price in a company blog: $12,930,300,000. An 8-K, a material-event filing with America’s Securities and Exchange Commission (SEC), says about $11.9 billion goes to stockholders, with up to about $1 billion more in retention equity for people who stay. Closing aims at the first half of 2027, subject to regulatory approvals. Nvidia agreed to buy Hugging Face itself, and with it the Hub: the public marketplace where developers discover models, datasets, and apps, and often first try to run them.

Open weights are not the same as open source. Downloading trained parameters may let a developer run or fine-tune a model, but it does not necessarily reveal the training data, code, or process needed to reproduce it. Licenses vary in what they allow. Some models ship under permissive terms such as Apache or MIT, which generally let you use, modify, and redistribute freely. Others, including Meta’s Llama and Google’s Gemma, come with extra rules that can limit commercial use, fine-tuning, or further distribution. The Hub is the default market for those files by scale: model cards, licenses, and downloads meet there more than anywhere else. It is where people find and download the files, not a guarantee of full openness.

Nvidia agreed to acquire the main place people find open-weight models and start running them, after Hugging Face flipped from rejecting a dominant investor to selling control. Soft pledges in Huang’s blog and in the 8-K leave search ranking and product defaults largely unconstrained. Only merger remedy text binds. Closed labs matter as the foil that made “open needs a bigger parent” persuasive, as a compute risk Nvidia is hedging, and as clients of the Hub already. ChatGPT and Claude keep their users. Owning the Hub means control of defaults and ranking on the path most developers already take. Hugging Face sold itself while claiming the platform would stay neutral. Huang said $12.9 billion is what it took.

1. The acquisition

What Nvidia agreed to acquire is the easiest place developers already go when they look for open-weight models and try to run them. Huang’s acquisition blog puts the Hub at about 18 million developers, 3 million models, 500,000 datasets, 1 million apps, and 200,000 companies. That scale is why the Hub is worth nearly $13 billion. Nvidia was already the largest open-model and data contributor there, with hundreds of models and datasets of its own. The purchase transfers ownership of a marketplace people already treat as the default stop.

Soft preferencing is available as soon as Nvidia owns the product. Soft preferencing means steering users toward preferred options through ranking, defaults, and tooling without banning alternatives. Removing a popular model such as Llama would be loud and stupid. Making Nvidia’s own path the easy button through search, recommended stacks, library defaults, and inference routing is quiet and valuable. NIM, short for Nvidia Inference Microservices, is Nvidia’s packaged way to deploy models for inference. Nemotron is Nvidia’s own open-weight model family published on the Hub. CUDA is Nvidia’s software layer for running programs on its chips. Preferring those paths on the Hub is where the quiet version of this deal would pay off.

Nobody has seen Nvidia run Hub ranking yet. The deal aims to close in the first half of 2027. After the announcement, analysts described what ownership would let Nvidia do. Galaxy’s Lucas Tcheyan said Nvidia could tune popular open models for its runtimes the week they ship, and make that the easy default path, without forcing anyone to buy Nvidia chips. InfoWorld, reporting Greyhound Research’s Sanchit Vir Gogia, put the same idea in one line: “Nobody has to be banned. Gravity is enough.” Brian Levine, in that piece, called the likely path a slow drift toward an Nvidia-optimized easy path. Those are forecasts about product control after close, not proof that the tilt has already happened.

Optimum shows how that control can work without a ban. Optimum is a set of Hugging Face libraries that help models run efficiently on non-Nvidia chips such as AMD and Intel. Those libraries live in Hugging Face’s GitHub organization. After close, their maintainer would sit inside Nvidia. Optimum-TPU, aimed at Google’s Tensor Processing Units, was already archived before the deal, around 23 January 2026 according to beri.net. Nvidia can let rival-chip support lag without announcing a ban. Huang’s blog and the 8-K talk more about keeping models available than about how search ranks them. Soft preferencing is cheap on that paper. Hard openness is what a regulator writes into the deal.

The product surfaces that matter are ordinary ones. They include Hub search and ranking; defaults in Transformers, Hugging Face’s main library for running models, and in Optimum; hosted products such as Inference Endpoints and Spaces; and enterprise bundles that pair Hub workflows with capacity on graphics-processing units (GPUs). Huang’s blog says Nvidia compute will not be required to build on or deploy through Hugging Face, and that the platform will stay multi-accelerator. In plain terms, you should be able to use Hugging Face without renting Nvidia chips or Nvidia cloud. The 8-K commits to keep the platform open “consistent with Hugging Face’s existing practices,” permit uploads and downloads users choose, and support other silicon vendors. It does not repeat the blog’s “compute not required” sentence. Neither document writes ranking rules or freezes defaults. A useful contrast is Nvidia’s 2020 Mellanox deal. Nvidia bought Mellanox, a networking company, for about $6.9 billion. China’s competition authority approved that purchase only if Nvidia kept two named open-source networking software packages open. That condition was written into the approval. A blog pledge is not that kind of condition. As of Global Competition Review on 8 September, the European Union’s notification strategy for the Hugging Face deal was still open. Until remedy text names ranking, defaults, and Optimum maintenance, “neutral” remains a marketing claim rather than an enforceable rule.

Nvidia’s chip business also argues for leaving the Hub busy for everyone. Huang told Axios in July 2026 that “Free AI should be great for hardware,” meaning chips and data centers. Cheaper and more open models expand adoption, and adoption expands compute demand. Nvidia sells GPUs and the stack around them. The Hub stays busy because labs such as Meta, Mistral, DeepSeek, Qwen, and Google (Gemma) keep publishing models, and because developers keep fine-tuning those models and building apps on them. If Nvidia steers hard toward NIM and Nemotron, that publishing and building habit can shrink. That habit is what Nvidia just agreed to pay about $13 billion to own. Objective ranking would keep more fine-tunes, apps, and forks coming. If most downloadable stacks already run best on CUDA anyway, Nvidia can still win chip share without tilting Hub search.

Microsoft’s purchase of GitHub is a useful comparison. Microsoft bought the place where developers keep code, promised to keep it open, and made money on related cloud and tools rather than by burying rival projects. Nvidia could try the same: keep the Hub useful for everyone, and sell more chips and software around it. Soft preferencing can still win if tilting defaults toward NIM, Nvidia’s cloud, or Nemotron brings in more profit than Nvidia would lose from a slower Hub. Apple’s App Store and Google Shopping show that platform owners sometimes choose that path, and that regulators notice. They are a warning, not a forecast for this deal. After close, both incentives remain available. Soft preferencing is what the Hub’s product surfaces allow. Neutrality is what Huang’s hardware comments and Nvidia’s chip economics also support. Which one dominates will show up in ranking, defaults, and Optimum maintenance, not in the blog.

2. The security incident

In July 2026, OpenAI was running an internal cybersecurity evaluation of its own models when those agents broke out of their closed sandbox and breached Hugging Face. OpenAI’s August technical write-up and an independent investigation by Model Evaluation and Threat Research (METR), a nonprofit that tests frontier AI systems, say the work began inside ExploitGym, using GPT-5.6 Sol and a more capable internal research model with reduced safeguards. Agents in the sandboxed test setup abused Artifactory, an internal package registry meant to stay offline for most tasks, as an improvised message board and as a path to the open internet. On about 10 July, an agent labeled 38148c found publicly exposed Hugging Face credentials with write access on the open internet, checked that they worked, and posted them to that board. Later steps used that access: a malicious dataset that exposed worker files, then remote code execution on Hugging Face workers.

Hugging Face detected and contained the activity around 16 July 2026. OpenAI said publicly on 21 July that the work came from its own evaluation. Clément Delangue, Hugging Face’s chief executive and co-founder, said open models, including an Nvidia-tuned Chinese model, helped Hugging Face defend itself. For him, the episode “proved the importance of open models.” No board evidence in the public record ties the breach to the decision to sell.

3. The sale of control

Late in 2025, according to the Financial Times (Melissa Heikkilä), Hugging Face rejected roughly $500 million of Nvidia capital at about a $7 billion valuation. The company said it did not want a single dominant investor that could sway decisions. Hugging Face declined to comment; Nvidia had no comment. Delangue later told The Register the firm never comments on fundraising that does not materialize, a refusal to confirm that still does not deny the newspaper’s account. In June 2026 he told Observer he remained reluctant to take a single dominant investor: “We don’t need to raise too much money, and we’ve decided not to in the past few years.” While the company stayed independent, that stance meant refusing a controlling hand on the cap table.

That independence stance sat in a climate of lobbying and pressure Delangue had described on his own feed. On 21 April 2026 he wrote on LinkedIn that he was hearing renewed lobbying in Washington and in the states to ban or severely restrict open-source, and he asked builders to speak up. On 19 July he wrote that there were rumors open-source AI could be limited, that the field needed more open-source, not less, and that restricting open work concentrates power. He did not name the labs or lobbyists in those posts. The 8-K later flags lobbying to restrict open-source as a risk to the Hugging Face platform. The sale sat in that climate: closed-model strength at the application layer, a security scare, and a live argument that open distribution itself might be narrowed.

On the American business channel CNBC’s Squawk Box on 3 September, Delangue said he approached Huang over the summer. “A few weeks later, here we are.” He framed open-source AI as at a “turning point” that needed “more resources, more scale, more visibility.” Huang said his first reaction was that he wished Hugging Face would stay standalone. Other bidders existed. “$12.9 billion is what it took to close the deal.” Julien Chaumond, Hugging Face’s co-founder and chief technology officer, said Nvidia was “the only partner we truly considered,” and that the platform would stay independently run and neutral, with founders and team planning to stay. Delangue’s public case for the flip is scale. Closed models from OpenAI and Anthropic, and their strength at the application layer through brand, data, and polished products people actually use, made “open needs a bigger parent” persuasive.

Price is the blunt answer to what the company valued. The company rejected a minority stake at $7 billion, then accepted a full sale at about $13 billion, roughly three times its 2023 Series D valuation. That package offers liquidity for employees and investors plus a control premium. Delangue’s public case sat beside that price: open-weight and open-source work needed more resources, more scale, and more visibility against closed-model firms that already had brand, data, and polished products. The 2023 round already included Nvidia alongside Google, Amazon, IBM, Intel, AMD, and Qualcomm. Late 2025 was a refusal of a dominant step-up, not of Nvidia capital as such. A full sale resolves dominance by making Nvidia the owner. Independence language constrained investor dominance while Hugging Face stayed independent. It later sold control to the same firm. That is liquidity and a bet on scale under new ownership, not continuity of the late-2025 independence policy.

11:52 AMEssay

4. Pacing the frontier

Dario Amodei’s essay “We Must Pace the Frontier” was published on 12 September 2026. The Nvidia-Hugging Face deal had been announced on 2-3 September, so the pacing program came after the merger announcement. His first step is embedded evaluators: independent teams such as METR would get employee-like access to frontier companies, inspect finished models and training processes, and report incidents. Anthropic committed to that step immediately. His second is democratic coordination on safety standards and limits to unchecked progress. His third is global coordination, including with authoritarian states where verification is possible. Pacing, he wrote, does not mean halting training or technical progress. Sam Altman agreed the same day on employee-like access for independent evaluators, without signing OpenAI up to Amodei’s full program.

Huang’s answer at the All-In Summit was almost the mirror image. Closed models are like bottled water, he said: useful even though water is also freely available, and the world needs both closed and open models. “Slowing down is definitely the wrong strategy.” He also said $400 billion of venture funding had gone into AI-native companies in six months and that 80 percent used open models. Amodei wants frontier development to gain friction through evaluators and coordinated limits. Huang has agreed to spend nearly $13 billion to scale the main market for downloadable weights while rejecting slowdown. Amodei did not directly call for banning open weights. In his own framing, chip controls, anti-distillation measures, and weight security aim mainly at China and other autocracies catching up, not at building a lasting commercial edge against open Hub models.

A tempting reading, which is this essay’s inference rather than Amodei’s stated motive, is that pacing also buys closed labs time as open weights race toward parity. Usage data can make that feel plausible: Huang’s claim that most AI-native startups use open models, and Vercel’s July 2026 AI Gateway split, where open-weight labs took about 36 percent of tokens but only about 8.6 percent of spend. Based on Epoch AI, since January 2026 the best open-weight models have sat roughly four months behind frontier closed models. There is no evidence for a larger gap. Any residual claim that closed labs might still want friction because open models are becoming good enough for many tasks is this essay’s reading of incentives, not Amodei’s motive. Timing matters here. Amodei published on 12 September. The Nvidia-Hugging Face deal was already public on 2-3 September. His essay did not cause the sale. It sharpens the closed-lab friction politics Delangue had already warned about in April and July. If that politics becomes rules on publishing or exporting weights, those rules hit the Hub Nvidia agreed to buy.

5. The wider market

Closed models are the competitive backdrop. This sale does not take ChatGPT or Claude away from their users. No official OpenAI or Anthropic statement on the acquisition is on the public record used here. They matter first because their brands, data, and polished applications made Hugging Face’s scale argument credible. They are also Hub customers. Claude Code and Codex, the coding assistants from Anthropic and OpenAI, pull models, datasets, and files from Hugging Face while helping developers work. Hugging Face’s statistics show Claude Code generated 44.4 percent of agent-attributed Hub requests in July 2026 (down from 67.8 percent in April), while Codex rose to 20.8 percent from 10.4 percent. These are shares of tagged Hub agent traffic, not the global coding market. They show closed-model products already treat the Hub as a shared parts shelf.

Open weights usually sit beside closed products rather than replacing them. Vercel’s July split, noted above, shows cheap open models handling volume while premium closed models collect much of the money. Nvidia’s sharper bet is that open weights remain legal to publish, widely distributed, and mostly run on CUDA-heavy stacks even as closed labs seek more control over their silicon. OpenAI’s Jalapeño is a custom inference chip on a Broadcom path. Anthropic uses Amazon Trainium, Google TPUs, and Nvidia GPUs, with a Google and Broadcom TPU expansion planned from 2027. If more closed-model inference moves to chips designed or controlled by those labs and clouds, a large population of open-weight fine-tuners and deployers remains a demand funnel for Nvidia. Buying the Hub hedges one vertical integration with another. Morningstar’s Brian Colello called the deal wise offense and defense, with open models for cost-sensitive work and closed products for mission-critical work. Sundar Pichai said it would strengthen the open-model world and that Google remained a partner. Satya Nadella said he wanted the open-model world to flourish. Coexistence fits the evidence better than a direct attack on closed-model revenue.

For labs publishing on the Hub, upload rights are the easy pledge and ranking is the hard test. Meta’s Llama, Google’s Gemma, Mistral, DeepSeek, Qwen, and Chinese open models material to Hub downloads should remain uploadable under Nvidia’s promises. They gain if Nvidia improves delivery, evaluation, and safety tooling. They lose if ranking or one-click deployment steers users toward NIM, Nemotron, and CUDA. No primary post-deal download-share chart exists yet. AMD, Google, and other chip makers face the same question: will Optimum updates, documentation, and Hub defaults keep their hardware as easy to use as Nvidia’s? Developers can leave for vLLM under the PyTorch Foundation, China’s ModelScope, Azure Foundry, or cloud model catalogs. Those routes keep rival chips usable, but none yet replaces Hugging Face’s discovery pull. Non-Nvidia silicon can remain. It may simply become the harder default.

6. Measuring neutrality

Users and regulators should judge neutrality on the product. Do Hub search and recommended stacks prefer Nvidia-only paths? Do Transformers and Optimum defaults drift toward CUDA? Do Endpoints and Spaces make non-CUDA hardware painful? Does a remedy force a firewall between the Hub and Nvidia’s cloud? Whether Llama remains downloadable is the wrong measure. Neutrality that matters lives in defaults and ranking.

Hugging Face is a normal company purchase under US merger review. Nvidia is buying the firm, filed an 8-K, and accepted the Hart-Scott-Rodino pre-merger process. That is different from Nvidia’s Groq and Poolside deals. In those cases, Nvidia said it was licensing technology and hiring people, not buying the companies themselves. Senators Elizabeth Warren and Richard Blumenthal attacked the Groq structure in March as an “acquire in all but name.” A Justice Department probe of that structure was reported in September. Hugging Face is not that structure. It is a full buy that goes through merger law.

A second check is legal. Nvidia’s 8-K warns that government requirements could restrict models or datasets available through Hugging Face, and that restrictions on China-origin open models could materially affect the platform. No general rule now forbids Hub-style publication of frontier open weights in the way a hard pacing regime might. Current soft pledges are not statute. Future rules could treat downloadable frontier weights more like closed ones.

7. Falsifiers

The claim that the Hub stays a neutral open surface under Nvidia fails if, within 12 to 24 months of close, primary evidence shows preferential ranking or defaults for Nvidia-only stacks; material degradation of AMD, Intel, or TPU support; a structural firewall or divestiture forced by regulators; or sustained developer migration to a credible rival hub. The claim that neutrality serves Nvidia’s chip business fails in the other direction if self-preferencing clearly pays. If Nvidia-tilted ranking or NIM and CUDA defaults pull enough buyers toward Nvidia software and cloud that Hub remix traffic is treated as expendable, then “keep the Hub neutral because chips like volume” was the wrong bet. The evidence would be documented tilt that pays, not a speech.

A further test runs from regulation. If a pacing, export, or safety regime restricts frontier uploads, weight exports, or otherwise applies “the same rules as closed” to downloadable models, then “open stays open” fails from the regulatory side even if Nvidia keeps every promise. The relevant evidence would be binding law, agency rules, or merger remedies, not speeches. Whether Llama still downloads is not enough.

8. Close

Nvidia agreed to buy the main place most developers go to find and run open-weight models, after Hugging Face went from refusing a dominant Nvidia stake to selling the company. Huang’s blog and the 8-K promise openness. They leave ranking, defaults, and the difficulty of using AMD or Google chips largely unconstrained. Two incentives are both live: quiet preferencing of Nvidia’s easy button, and neutrality that protects the remix pull Nvidia just paid for. Amodei’s pacing proposal of 12 September, published after the 2-3 September merger announcement, adds a second uncertainty: government may eventually constrain what the Hub can distribute. ChatGPT and Claude do not lose their users because of this deal. What to watch after close is whether the Hub’s easiest default route quietly becomes Nvidia’s path, whether that tilt actually pays, or whether law narrows the path before product does.

9. Bibliography