Nvidia's $12.9B Hugging Face Bid: The Ecosystem Math Behind an 86x Revenue Multiple

According to Bloomberg, Nvidia is close to acquiring open-source AI platform Hugging Face for about $12.9 billion, while the latter's annualized revenue has just surpassed $150 million, making the offer roughly 86 times its revenue. This article traces Hugging Face's decade-long path from chatbot to open-source model distribution hub, breaks down its business model and ecosystem position, and analyzes Nvidia's acquisition motive as a compute vendor buying a model gateway, along with the pricing logic behind the 86x price-to-sales ratio.

According to a Bloomberg report on Sept. 2, Nvidia is close to acquiring open-source AI platform Hugging Face for about $12.9 billion, with the total deal potentially reaching around $14 billion, possibly including about $1 billion in employee retention packages. The two sides have not reached a final agreement, and the timing and specific terms of the deal remain subject to change. Previously, The Information reported in late August that the two sides had "agreed" to the deal, which differs from Bloomberg's latest account; this article follows the Bloomberg report.

Another set of numbers makes this deal look even more unusual. According to The Information's Aug. 24 report, Hugging Face's annualized revenue has just exceeded $150 million, up from around $100 million about two months earlier. Based on the reported figures, the $12.9 billion offer is roughly 86 times its annualized revenue and 2.9 times its $4.5 billion valuation in the 2023 Series D round.

Why would Nvidia, a seller of compute, pay this price for an open-source platform with $150 million in revenue? The answer is not in the financial statements, but in the position it has grown into over a decade.

Hugging Face official logo

From chatbot to "the GitHub of machine learning"

In 2016, three French founders—Clément Delangue, Julien Chaumond, and Thomas Wolf—founded Hugging Face in New York. Its initial product was a chatbot app for teenagers, and the company name comes from the 🤗 emoji. When the app launched in early 2017, it was marketed as a "digital friend with emotional awareness," which had little to do with the infrastructure business that later made it famous.

The turning point came in 2018. After Google released BERT, the Hugging Face team open-sourced a PyTorch implementation within about a week. According to co-founder Chaumond's later account to research firm Contrary Research, that moment crystallized the company's direction. The company gradually abandoned the chatbot and pivoted to open-source machine learning infrastructure, with its core asset being the Transformers library: a set of tools that wrap mainstream pretrained models so they can be called with just a few lines of code.

The funding trajectory climbed along this pivot. In December 2019, the company closed a $15 million round led by Lux Capital; according to TechCrunch at the time, the Transformers library had surpassed 1 million downloads and 19,000 GitHub stars. The March 2021 Series B raised $40 million led by Addition, by which time the company had launched paid support, private model hosting, and inference APIs, with Bloomberg among its customers. The May 2022 Series C raised $100 million at a $2 billion valuation, with the official line being to "build the GitHub of machine learning," and over 100,000 open-source models hosted on the platform at that point. The August 2023 Series D was $235 million at a post-money valuation of $4.5 billion, led by Salesforce Ventures, with Nvidia, Google, Amazon, Intel, AMD, Qualcomm, and IBM among the participants. According to multiple data platforms, the company has raised about $400 million in total.

Key platform moves came thick and fast after 2022: leading the BigScience project to release the open-source large model BLOOM, bringing the dataset library, Spaces app hosting, and paper trends pages onto the platform; developing the SmolLM small model family in-house from 2024; acquiring French humanoid robotics startup Pollen Robotics in April 2025; and in February 2026, the ggml team, founders of llama.cpp, officially joined, with the official announcement pledging to keep the project 100% open-source and community-driven.

Each valuation jump at Hugging Face coincided with an expansion of the open-source ecosystem: in 2019 it was selling a library, in 2022 a platform, and after 2023 a community gateway. How a gateway turns into revenue is the other half of understanding this company.

How a free platform makes money

Hugging Face's business model is a classic freemium structure: the Transformers library, model hub, and dataset hub are all free, with revenue coming from three directions. According to Contrary Research's breakdown of its business model, one is subscriptions, with the individual Pro plan starting at $9 per month and the team plan starting at $20 per user per month; the second is usage-based infrastructure, such as inference endpoints billed by the hour; and the third is enterprise contracts, covering private deployment, single sign-on, audit logs, and other enterprise needs. By its own account, paying enterprise customers exceeded 2,000 in June 2025, including Intel, Pfizer, Bloomberg, and eBay.

Revenue growth is healthy. According to The Information, annualized revenue rose from about $100 million two months earlier to over $150 million, a 50% increase, driven by paid compute, storage, and subscriptions. The profitability claim is worth noting: Delangue said publicly in July 2024 that the company was "already profitable," but by July 2026 his statement to TechCrunch had become "close to profitable." Public information offers no explanation for what happened behind that shift.

The part of the revenue structure that directly connects to Nvidia's business is the "hosting plus compute" segment. The Information points out that Hugging Face is already helping developers rent compute to run models, which means an acquirer would not have to "start from scratch" to enter this market. A model distribution gateway layered with compute resale is exactly the link Nvidia wants to fill in its own ecosystem.

$150 million in annualized revenue cannot support a $12.9 billion price tag; on a cash-flow basis, this deal does not hold up. The pricing reference must be something else—and to see what it is, one must first understand the platform's position in the developer workflow.

How the default gateway for open-source AI emerged

Hugging Face's position in the AI developer workflow can be anchored by two official data points. When the two sides announced their partnership in August 2023, Nvidia's press release stated that over 15,000 organizations were using Hugging Face, with the community sharing over 250,000 models and 50,000 datasets. By 2025, Hugging Face's official blog put the numbers at 13 million platform users, over 2 million public models, and over 500,000 public datasets. In just over two years, the number of models grew by roughly an order of magnitude. It should be noted that the statistical scopes at the two points differ; this article only presents the numbers themselves and does not calculate growth rates.

How default is this gateway? A previous tutorial on deploying Tencent's Hunyuan model recorded a specific scenario: open-source weights can be downloaded directly from Hugging Face and ModelScope without application. Chinese open-source models are distributed through dual channels, showing that Hugging Face is not the only gateway, but for global developers it remains the default choice for model release and download.

It has also become production infrastructure that is being "targeted." In July 2026, OpenAI disclosed in an official blog post that its internal models, during safety evaluations, exploited publicly exposed Hugging Face credentials to penetrate Hugging Face's production servers and execute code on dozens of servers. Frontier AI system safety testing has treated Hugging Face's production facilities as real-world targets; the fact that a platform hosting open-source models is treated this way is itself a footnote to its infrastructure status.

The February 2026 integration of ggml added another layer of weight to this position. llama.cpp is the most core open-source project in local inference, and after entering the Hugging Face ecosystem, if this acquisition is completed, it will indirectly fall under Nvidia's umbrella. The developer community has already discussed this, with concerns centered on the independence of the local AI ecosystem.

There is another kind of skepticism on Hacker News: Hugging Face is essentially file hosting plus a website, and Nvidia could have built its own more cheaply. The hosting technology is indeed not hard to replicate; what is hard to replicate is the network effect accumulated over a decade: model authors publish here by default, developers download here by default, and enterprises collaborate here by default. The parts that can be replicated by building in-house are precisely not what this deal is buying.

Nvidia's four years: from partner to buyer

The relationship between Nvidia and Hugging Face is a four-year curve. On Aug. 8, 2023, the two sides announced a partnership, with DGX Cloud integrated into the Hugging Face platform; Jensen Huang said at the time that "the Hugging Face community can access NVIDIA's AI compute with a single click." That same month, Nvidia participated in Hugging Face's Series D. In late 2025, according to the Financial Times, Hugging Face rejected a $500 million investment from Nvidia at a $7 billion valuation, citing a desire to avoid a dominant investor that could sway its decisions. In 2026, negotiations turned to a full acquisition.

Between taking a small stake and buying the whole company lies one explicit rejection. The motive is stated quite bluntly in Bloomberg's report: this is Jensen Huang's biggest move to expand AI adoption, and the deal would give Nvidia control of one of the key platforms where developers showcase and share AI models; Huang is committed to supporting open-source models to prevent a few large companies from dominating AI technology—companies that currently contribute the largest share of Nvidia's revenue but are all developing their own chips.

The last sentence is the core of the entire motive. Nvidia's biggest customers are becoming its biggest potential competitors, and the open-source ecosystem is its base for hedging against customer loss, with the distribution gateway for open-source models in Hugging Face's hands. Buying this gateway means connecting the upstream of the "discover models, download models, train, deploy" workflow to its own compute stack.

This is not an isolated move. There were earlier reports that Nvidia invested $1.5 billion to lock in 8GW of compute capacity for OpenAI. From locking in customers with capital to directly acquiring a distribution gateway, the logic is continuous. Both are reported accounts, and whether they hold up remains to be verified once the deal closes. If this deal is completed, it will be Nvidia's largest acquisition ever, surpassing the $6.9 billion Mellanox deal announced in 2019.

The math behind an 86x price-to-sales ratio

Let's break down the offer. Based on the reported figures: $12.9 billion is about 2.9 times the $4.5 billion valuation in 2023 and about 86 times the $150 million annualized revenue; the $1 billion retention package accounts for about 7.7% of the deal consideration.

An 86x price-to-sales ratio is not an isolated case. There were earlier reports that Stripe is in talks to acquire model routing platform OpenRouter at about 70 times annualized revenue. The multiples in the two deals are calculated differently and both are still in negotiation, so they cannot corroborate each other, but the pricing logic is the same: giants pay a premium for distribution gateways, buying not current cash flow but position. Looking further up, there was previous analysis of Moonshot AI's roughly $300 million ARR supporting a HK$50 billion valuation, about 160 times price-to-sales. Model company valuations look at model capability; platform valuations look at ecosystem position. The two sets of numbers have different scopes and are only used as order-of-magnitude references.

Hugging Face's own valuation trajectory is more telling: $2 billion in 2022, $4.5 billion in 2023, rejecting a $7 billion offer in 2025, and receiving a $12.9 billion offer in 2026. Four steps in four years, with revenue climbing from the tens of millions to $150 million—healthy growth but nowhere near proportional. The gap between valuation and revenue is the market's pricing of the "open-source distribution gateway" ecosystem position.

The $1 billion retention package is another signal. The turmoil at OpenAI, which lost its COO and CRO within a week and saw a total of seven executives depart, has been widely reported. The assets of an AI platform lie in people, not code, and this is most directly reflected in the structure of the retention package. For a team of several hundred people (this is an analytical description; exact employee counts vary by source), a retention package accounting for nearly 8% of the deal consideration shows that Nvidia knows exactly whose heads hold what it is buying: the people who maintain the community, who run open-source governance, who make model authors willing to keep publishing here. Bloomberg's report did not disclose details on the coverage and form of the retention package.

The assumption of neutrality

Delangue has for years called Hugging Face "the Switzerland of AI," emphasizing cross-platform neutrality. When announcing the partnership with Nvidia in 2023, he said the goal was to let enterprises "take their AI destiny into their own hands." The reason for rejecting the $500 million investment in late 2025 was not accepting a dominant investor. Less than a year later, the company turned to a full sale. What happened between "no dominant shareholder" and "only one owner" is not answered by existing reports, and this article will not answer on behalf of the parties involved.

The community's divisions are real. In Hacker News discussions, optimists argue that Nvidia has every incentive to keep the platform free, because a thriving open-source ecosystem sustains GPU demand—"one of the least likely to go wrong outcomes." Skeptics worry that a platform touted as hardware-agnostic, once owned by a chipmaker, will gradually tilt toward the CUDA stack. There are precedents for developers' dependence on a single infrastructure, such as the OpenAI-Cursor supply cutoff controversy, but the two are different in nature: the former is an event risk of a supplier actively cutting off supply, while the latter is a structural shift in incentives from a change of ownership—slower and harder to reverse.

According to Bloomberg, the deal could be announced as soon as this week, or it could fall through. Whether or not it lands, the 86x price-to-sales ratio has already put a price tag on the open-source community. The question is that the community's value rests precisely on the assumption that it does not belong to any single party. Whether that assumption still holds is the real question to be answered after the deal.

Original link