Nvidia’s Hugging Face Deal Secures the Model Choice Layer

Nvidia has agreed to buy Hugging Face for $12.9 billion, The Information reported on August 26, citing a person familiar with the matter. The price sits near 86 times the open-model hub’s roughly $150 million annualized revenue and nearly triples its $4.5 billion valuation from the 2023 Series D.

Business Insider described talks at a valuation above $13 billion that had not produced a signed agreement and could still collapse. Neither company has confirmed the reports. The deal, if completed, would give the dominant AI chipmaker ownership of the platform where developers discover, download and deploy the open weights that keep demand flowing to its GPUs.

The gap between the two accounts matters. One describes an agreed price. The other stresses that nothing is signed and that the talks remain fragile. Until a joint statement or a filing appears, the market is left reading secondhand reports.

Nvidia Locks the Model Choice Layer

Closed-source labs including OpenAI, Anthropic and Google are designing their own accelerators to reduce reliance on Nvidia silicon. A thriving open-weights ecosystem spreads compute across hundreds of thousands of developers who buy retail hardware and rarely negotiate. Owning the hub that hosts those models lets Nvidia steer that traffic.

Hugging Face already helps users run models on rented compute. TechCrunch noted that control of the platform could let Nvidia recycle unused cloud capacity it has committed to customers. The same ownership also positions the company for a return to cloud services after it scaled back its DGX Cloud effort.

That dual use is straightforward. Developers already come to the hub to find weights and spin up demos. Routing a share of that work onto idle committed capacity turns a fixed cloud obligation into a live product surface. It also softens the earlier retreat from DGX Cloud by giving Nvidia a consumer-facing front door again.

The reported price equals about six weeks of Nvidia’s recent sales pace. The company posted $96.2 billion of quarterly revenue for the period ended July 26, up 106 percent year over year.

At that run rate, a $12.9 billion check is large in absolute terms and modest against the cash the chipmaker is generating. The strategic question is not whether Nvidia can afford the hub. It is whether ownership of the discovery layer is worth paying a premium that far above trailing sales.

From $4.5 Billion Round to $12.9 Billion Report

Hugging Face raised $235 million in August 2023 at a $4.5 billion valuation. Salesforce Ventures led; Nvidia, Alphabet’s GV, Amazon, IBM, Intel, AMD, Qualcomm and others participated. The company has not raised externally since.

Late last year it rejected a $500 million Nvidia investment that would have valued it at $7 billion. The Financial Times reported the decision rested on a desire to avoid any single dominant backer that could sway decisions at a platform meant to stay neutral.

Co-founder and CEO Clément Delangue has said the company was close to profitability and had only recently begun spending the 2023 capital. Co-founder Thomas Wolf, asked about the sale reports on Bloomberg Tech, declined to comment on Nvidia specifically but said Hugging Face has “always had a lot of interest.”

• $4.5 billion, 2023 Series D post-money valuation
• $7 billion, rejected Nvidia investment valuation, late 2025
• $12.9 billion, reported agreed price, August 2026
• ~$150 million, recent annualized revenue run rate

1. August 2023: Series D closes at a $4.5 billion post-money valuation with a broad investor roster that already included Nvidia.
2. Late 2025: A $500 million Nvidia primary investment at a $7 billion valuation is rejected on governance grounds.
3. August 2026: The Information reports an agreed full acquisition at $12.9 billion; Business Insider still describes unsigned talks above $13 billion.

At the reported figure the multiple lands near 86 times sales. Sacra estimates the platform reached $150 million in annualized revenue in August 2026 after hitting $100 million only two months earlier.

The jump from $100 million to $150 million in two months shows how fast the top of the funnel is compounding. Even so, an 86 times sales multiple prices the hub as infrastructure, not as a conventional software business. The earlier $7 billion conversation already implied a rich multiple. The reported buyout nearly doubles that mark again.

What the Hub Holds

Hugging Face’s own summer 2026 State of Open Models report supplies the inventory numbers. Public model repositories grew from 2.43 million to 2.96 million public model repositories between January and August. Datasets rose from 711,000 to 1 million. Hosted Spaces (demo applications) climbed from 1 million to 1.44 million.

Concentration is extreme. Roughly 85.6 percent of repositories have been downloaded fewer than 200 times. Just 1.5 percent of repositories account for 99.2 percent of all downloads. Qwen derivatives from Alibaba now number 151,448 repositories, 2.6 times Meta’s footprint, and are added at 180 to 210 per day. Only about 3 percent of 2026 downloads went to models above 70 billion parameters; small models remain the practical workhorses.

Chinese labs have set the monthly frontier ceiling for open models in almost every month of 2026. Nvidia and AMD themselves rank among the most prolific publishers of new open model repositories this year, each exceeding 200.

The inventory growth and the download skew point in the same direction. The hub is vast at the long tail and narrow at the head. Whoever controls placement, defaults, and runtime hints around that top 1.5 percent shapes most of the traffic that actually moves silicon.

Small models dominate practical use. That pattern favors the retail and prosumer GPU buyers who rarely negotiate enterprise contracts. It is the same population the open-weights story was always meant to serve, and the same population Nvidia reaches most directly through its existing channels.

The Independence Stance That Flipped

Delangue spent 2026 publicly aligned with Nvidia’s defense of open weights. He appeared on CBS Face the Nation and CNBC citing a letter signed by Jensen Huang and two dozen companies, including Hugging Face, that urged Washington against premature restrictions. Hugging Face itself used an Nvidia-modified Chinese open model to investigate a security incident after closed models’ guardrails blocked forensic work.

The earlier rejection of Nvidia capital rested on governance. A full acquisition removes the dominant-investor problem by transferring ownership outright. Some observers see that distinction as the reason a buyout could succeed where an oversized stake failed. Others note the complete reversal of the long-term sustainability language Delangue had used when he said the company was not optimizing for an exit.

Founded in 2016 in New York by French entrepreneurs Delangue, Julien Chaumond and Thomas Wolf, the company began as a chatbot app, open-sourced its transformers library, and became the default warehouse for open models and datasets. It still describes itself as community infrastructure.

The public record now holds both positions in plain view. Delangue argued for open weights beside Nvidia’s chief executive, then watched his company reject a primary check from the same firm, and now faces reports of a full sale to it. The governance logic of a clean sale differs from the logic of a blocking stake. The optics of the path still reverse years of independence language in a single step.

Developers Watch for CUDA Bias

Reaction on X split quickly. Some accounts welcomed Jensen Huang’s record of supporting open models. Others warned that the chip vendor would favor CUDA paths, de-emphasize AMD, Intel or Apple MLX runtimes, and turn the hub into a sponsored channel. One widely shared take argued that open weights had become a distribution play and that a torrent-style alternative might eventually be needed.

The value of the platform rests on perceived neutrality. Developers arrive because the hub does not appear to sell anything. If non-Nvidia runtimes degrade or sponsored models receive preferential placement, traffic can leave. Forbes contributor Jon Markman framed the test simply: watch whether AMD, Intel and hyperscaler paths still work as well twelve months later. If they do, Nvidia bought a public road. If not, it bought an archive.

Nvidia buying Hugging Face is not a win for open source. They don’t open CUDA. They don’t OSS drivers. Now they want Hugging Face. They will push NVIDIA-sponsored models, cap downloads or may charge for training and starve Apple MLX, AMD, Intel anything not CUDA.

prayag sonar, @prayag_sonar on X

Crowd intelligence also noted the parallel to CUDA itself: free software that became the installed-base moat. The hub could function the same way for model selection.

That parallel is the core risk developers are pricing. CUDA won by being free, fast, and everywhere, then by becoming hard to leave. A model hub that stays open on paper while steering defaults, containers, and one-click paths toward one vendor’s stack would repeat the pattern at the selection layer rather than the kernel layer.

Markman’s twelve-month test is measurable. If AMD, Intel, and Apple MLX paths keep working without friction, the neutrality claim holds. If they quietly worsen, the archive reading wins.

Regulators and Rivals Sit Quiet for Now

No formal filings have appeared. Antitrust review would be required in multiple jurisdictions given Nvidia’s market position and the hub’s role as essential infrastructure for open AI. Chinese model families already dominate downloads and derivatives; an American chipmaker owning the main Western distribution point will draw attention in Beijing and Brussels alike.

Rivals that publish on the hub-Meta, Google, Mistral, Alibaba, Moonshot-face a new gatekeeper. Closed labs building custom silicon lose one more independent distribution channel. Hardware competitors lose a neutral showcase. Hugging Face employees and the 13 million registered users gain deeper pockets and tighter integration with the dominant accelerator stack, at the cost of independence.

Stripe’s recent acquisition of OpenRouter for more than $7 billion, a company valued at $1.3 billion months earlier, shows that distribution layers in AI are clearing at elevated multiples. Hugging Face is larger and more central.

Both deals pay up for routing layers rather than for model training labs. OpenRouter was smaller and newer. Hugging Face already concentrates discovery, downloads, datasets, and demo Spaces in one place, which helps explain why the reported check is higher still.

Ownership Changes the Neutrality Bargain

The hub’s brand was built on the claim that it did not belong to any single chip vendor or closed lab. That claim attracted the long tail of publishers and the concentrated head of downloads alike. A completed sale replaces the claim with a different bargain: deeper capital and tighter GPU integration in exchange for an owner whose main business is selling the accelerators underneath.

The rejected $500 million primary stake failed because a dominant minority investor could tilt product choices while the company still called itself neutral. A full acquisition ends that half-measure. It also ends the ability to say the platform has no owner with a hardware agenda.

Developers, rival chip firms, and Chinese model labs will all read the same shift. Their responses will show up in runtime defaults, mirror traffic, and any move toward torrent-style distribution of the kind already floated on X. Those responses, more than the closing documents, will decide whether the hub remains a live road.

What Still Has to Happen Next

As of August 29 the reports remain unconfirmed by either party. Business Insider continues to stress that talks can still fall apart. The Information describes an agreed deal. The next concrete step will be a joint announcement, a regulatory filing, or a denial.

Until one of those arrives, the story sits on two competing wires and a set of public facts that already explain the stakes. The multiple, the inventory concentration, the prior rejection of Nvidia capital, and the CUDA bias debate are all on the record. Only the signature and the regulators are not.