Signals Inbox·September 2, 2026·AI Infrastructure

Why is Nvidia buying Hugging Face?

Nvidia wants Hugging Face because the next AI fight is moving upstream from chips to the route between models and compute. Hugging Face gives it distribution, developer habits, open-model infrastructure and a foothold in robotics and local AI, provided Nvidia does not destroy the neutrality that made the platform valuable.

Q1Is Nvidia actually buying Hugging Face right now?

As of now, Nvidia is clearly trying to buy Hugging Face, but the freshest reporting says the deal still has not reached a final agreement.

The status is messier than the headlines make it look. The Information reported that Nvidia had agreed to acquire Hugging Face for $12.9 billion. The freshest Bloomberg reporting, based on people familiar with the talks, says the two sides are still in advanced negotiations and that a final agreement has not been reached. Bloomberg also says the package could include roughly $1 billion to retain Hugging Face employees, which would push the total cost close to $14 billion.

We should therefore treat the acquisition as highly advanced rather than completed. Nvidia and Hugging Face have both declined to publicly confirm the terms, and there is no announced closing. If the deal lands around the numbers now being discussed, it would also become Nvidia's largest acquisition, well above the roughly $6.9 billion it paid for Mellanox.

Current acquisition picture

Q2Why does Nvidia want Hugging Face now?

Nvidia wants Hugging Face now because open AI has become large enough to influence where a serious amount of future compute gets spent.

What changed lately is the business side. The Information reported that Hugging Face's annualized revenue jumped from roughly $100 million to more than $150 million in only two months. That figure extrapolates the latest monthly revenue rather than representing contracted ARR, but a 50% jump in that short a period is still hard to ignore. CEO Clément Delangue also said the number of paying subscribers doubled during the first half of 2026.

Usage is moving in the same direction. Hugging Face's summer ecosystem report shows public model repositories rising from 2.43 million to 2.96 million during the first seven months of 2026, while datasets grew from 711,000 to one million and Spaces from 1.00 million to 1.44 million. That works out to roughly 22% growth in model repositories, 41% in datasets and 44% in Spaces in a little over half a year.

The platform is adding content much faster while finally turning more of that activity into paid compute, storage and enterprise subscriptions. Nvidia has been close to Hugging Face for years. The case for owning it looks much stronger now than when Hugging Face was mainly seen as an open-model community.

Q3What is Nvidia actually buying with Hugging Face?

Nvidia is trying to buy the main distribution layer for open AI: the place where developers discover models, modify them, share them and increasingly pay to run them.

Hugging Face now sits across several steps of the AI workflow. The Hub hosts models and datasets, Transformers is one of the standard libraries used to work with them, Spaces lets developers publish applications, Inference Endpoints sells managed deployment, and the enterprise product lets companies keep models and data private.

The audience behind those products is already huge. Hugging Face said it reached 13 million users in 2025, and a company executive said at Nvidia GTC that more than 300,000 organizations use the platform. He also said private organizational models were roughly as numerous as public ones at that point. The private-model figure shows how much Hugging Face activity sits behind company firewalls rather than on the visible public Hub.

The value comes from where Hugging Face sits in the decision chain. Developers often choose a model, library and deployment workflow before they choose the hardware that will eventually run the workload. Hugging Face gives Nvidia a place much closer to that first choice.

Q4Is Hugging Face really dominant enough to justify Nvidia's takeover bid?

Hugging Face is dominant enough for Nvidia to pay up because recreating its scale and developer habits from scratch would be painfully difficult.

The public Hub passed one million models in 2024, crossed two million during 2025 and moved above three million this year. That pace shows how deeply the platform has become embedded in the way open models get published and reused.

The interesting part is how concentrated real usage remains. Hugging Face's latest ecosystem analysis found that 85.6% of model repositories have fewer than 200 lifetime downloads, while just 1.5% of repositories account for 99.2% of all downloads. A relatively small set of models therefore drives almost all consumption, but those winners are discovered, forked, benchmarked and distributed inside the same enormous network.

For Nvidia, that concentration is actually useful. Only a small group of models has to become commercially important for Hugging Face to sit in the middle of their spread through developers and companies.

Hugging Face's scale and usage concentration

Q5Why does open AI make Hugging Face so valuable to Nvidia?

Open AI makes Hugging Face valuable to Nvidia because a market full of independent model builders is safer for its chip business than a market controlled by a few giant labs with the scale to fund alternative silicon.

We can already see that pressure. OpenAI unveiled its own inference accelerator with Broadcom this year. Anthropic says it now uses more than one million Amazon Trainium2 chips and has committed more than $100 billion over ten years to AWS technologies. Google keeps expanding its TPU line, while Amazon says Trainium and Graviton together have passed a $10 billion annual revenue run rate.

Nvidia is still growing at extraordinary speed. Its latest quarter produced $89 billion of Data Center revenue, up 117% from a year earlier. Yet Nvidia's SEC filing also shows how concentrated the business remains: three direct customers accounted for 16%, 15% and 13% of total revenue during the first half of the fiscal year. Together, that is 44%.

That combination explains Nvidia's open-model push better than any slogan about openness. Nvidia launched the Nemotron Coalition with Mistral AI, Perplexity, Cursor, LangChain, Thinking Machines Lab and others, and the first coalition model is being trained on Nvidia DGX Cloud. The more model creation spreads across many independent companies, the harder it becomes for a handful of vertically integrated labs or cloud providers to dictate the hardware underneath the whole market.

Q6How would owning Hugging Face actually help Nvidia sell more GPUs?

Owning Hugging Face would give Nvidia a powerful new way to sell more GPUs by putting it closer to the moment developers choose where an open model will run.

Hugging Face already sells managed inference across several hardware options. Its current Inference Endpoints catalog includes Nvidia T4, L4, A10G, L40S, A100, H100 and H200 instances, alongside AWS Inferentia2 and Google TPU capacity. Nvidia would therefore be acquiring a platform where developers already compare and consume competing compute.

Small technical defaults are where this gets interesting. Nvidia can make sure major models receive CUDA, TensorRT and NIM optimization quickly; publish deployment recipes earlier; tighten integration between Transformers and Nvidia software; and make Nvidia-backed endpoints easy to use. None of this requires Hugging Face to shut out another chipmaker. A smoother path can shift a lot of workloads when millions of developers repeat the same choices.

Nvidia also has a very current reason to want more traffic flowing into compute. Nvidia's latest SEC filing says the company has $36 billion of multi-year cloud commitments with AI cloud partners. Those commitments fall as third-party customers consume the capacity. Hugging Face already has developers arriving with a model they want to run, so it could become one of the cleanest ways to send real customers into Nvidia-backed infrastructure.

Q7Can Nvidia own Hugging Face without ruining its neutrality?

Nvidia can preserve Hugging Face's value only if developers and competing chipmakers still believe the platform gives them a fair place to build and deploy AI.

Hugging Face was deliberately built across rival ecosystems. Its 2023 funding round included Nvidia, Google, Amazon, AMD, Intel, Qualcomm, IBM and Salesforce. The platform currently exposes Nvidia GPUs, Google TPUs and Amazon Inferentia through its own deployment products. That mix helped Hugging Face become a neutral meeting point rather than another vendor-specific developer portal.

We saw the neutrality issue before the takeover talks. The Financial Times reported that Hugging Face rejected a $500 million Nvidia investment that would have valued the company at roughly $7 billion because management wanted to avoid giving one dominant investor too much influence. A full acquisition would obviously create far more control than the investment Hugging Face previously resisted.

Nvidia has to be careful here. Favoring its own hardware quietly through better optimization is commercially rational. Making Hugging Face feel like an Nvidia storefront could push AMD, AWS, Google and model developers to invest harder in alternatives. Much of the underlying ecosystem is open source, which makes that threat credible: libraries can be forked, models can be mirrored, and communities can move.

Q8Why can't Nvidia just build its own Hugging Face?

Nvidia can copy most Hugging Face product features, but it cannot cheaply copy years of developer habit, links, integrations and community history.

Nvidia already operates NGC, a catalog for GPU-optimized models, containers, SDKs and deployment software. Technically, Nvidia has more than enough engineering talent to build another repository, another model browser and another managed inference product.

The harder problem is getting everybody else to show up. When a new open model launches, Hugging Face is already one of the default places developers check. Fine-tunes appear there, evaluation results point there, tutorials link there, libraries integrate with it and companies create private organizations around the same workflow. Each layer reinforces the others.

Microsoft's GitHub acquisition is a useful comparison. Microsoft could have built another code-hosting platform in 2018, yet paying $7.5 billion for GitHub gave it an existing developer network of roughly 28 million people. By 2022, GitHub had more than 90 million active users and about $1 billion of ARR. The code is the easy part here. Rebuilding the habit would be much harder.

Q9Is robotics a hidden reason Nvidia wants Hugging Face?

Robotics is one of the strongest hidden reasons Nvidia wants Hugging Face today.

Nvidia itself says robotics is the fastest-growing category on Hugging Face and that Nvidia's open robotics models and datasets lead downloads in that category. The two companies have also moved from loose compatibility to direct product integration. Nvidia's Isaac GR00T 1.7 and Isaac Teleop are now available through Hugging Face's LeRobot framework, and Nvidia says Cosmos 3 integration is planned next.

The overlap is huge. Nvidia says it has around two million robotics developers, while Hugging Face brings the 13 million AI builders mentioned earlier. LeRobot also reaches developers working with actual machines, including humanoids and low-cost robot arms, while Nvidia wants Jetson, Isaac, GR00T and Cosmos to become common infrastructure for physical AI.

If robotics becomes a major new compute market, Nvidia wants the model and dataset workflow to feel natural on Nvidia hardware before the market settles on its long-term software standards. Hugging Face gives Nvidia an existing community where those standards are already being shaped.

Q10Why does llama.cpp matter in Nvidia's Hugging Face deal?

llama.cpp makes Hugging Face more valuable to Nvidia because it gives the platform a foothold in local AI as some workloads move away from giant cloud GPU clusters.

The GGML team behind llama.cpp joined Hugging Face earlier this year, while keeping the project open source, community governed and technically autonomous. Hugging Face called llama.cpp a fundamental building block for local inference and said the goal was to make new models flow from Transformers into llama.cpp with much less friction.

The usage data backs up the importance of that layer. Among Hugging Face models that declare a parameter count, models below one billion parameters account for 83% of all-time downloads. Models above 100 billion account for only about 1%, and even during 2026 models above 70 billion represented just 3% of download volume.

Local AI can reduce the amount of inference that runs in hyperscale clouds, but Nvidia still has plenty of hardware on the receiving end: RTX GPUs, workstations, DGX Spark and Jetson systems. For Nvidia, the useful part is being close to the software standards that decide how efficiently open models run on all kinds of local devices.

Q11Is Nvidia overpaying for Hugging Face?

As seen above, Hugging Face is extremely expensive on normal software math: $12.9 billion against more than $150 million of annualized revenue implies a valuation of roughly 86 times revenue.

The jump is also large against Hugging Face's own history. Salesforce Ventures led a 2023 round at a $4.5 billion valuation, with Nvidia among the investors. The later Nvidia proposal reported by the Financial Times valued Hugging Face at about $7 billion. The takeover price now being discussed is roughly 2.9 times the 2023 valuation and about 84% above that rejected $7 billion proposal.

Recent developer-platform deals show that strategic buyers are willing to pay huge multiples, but Hugging Face still sits at the expensive end. Stripe's OpenRouter acquisition was reported around $7.5 billion to $8 billion after OpenRouter reached roughly $140 million to $160 million in annualized revenue. Microsoft paid $7.5 billion for GitHub when GitHub was reportedly producing about $200 million to $300 million of ARR.

The price is easier for Nvidia to absorb than those multiples suggest. Nvidia's latest quarter produced $59.7 billion of net income, and its SEC filing showed $56.6 billion of cash, cash equivalents and marketable debt securities. The reported purchase price would equal roughly 22% of one quarter's profit and 23% of those liquid holdings.

Developer-platform acquisition multiples

Q12Could regulators block Nvidia from buying Hugging Face?

Yes, regulators could block Nvidia's Hugging Face deal, and the failed Arm acquisition gives them a ready-made theory for doing it.

The FTC sued to block Nvidia's $40 billion Arm acquisition in 2021. Its argument focused on Arm's role as a neutral supplier whose customers included Nvidia competitors. The agency said Nvidia ownership could give the company access to sensitive rival information, weaken Arm's incentive to support technologies that competed with Nvidia and damage the trust that made Arm useful across the chip industry.

Hugging Face creates a similar question at the AI software layer. AMD, Amazon, Google and other Nvidia competitors integrate with the platform. Model developers also publish, benchmark and sometimes privately store important work there. A regulator could reasonably ask whether Nvidia would have the ability or incentive to favor its own hardware, use commercially sensitive information or make rival accelerators less attractive.

Arm and Hugging Face differ in one big way. Hugging Face contains far more open-source software and portable model artifacts, so developers have more realistic ways to fork projects or move elsewhere. That makes it harder for Nvidia to shut rivals out, and regulators could demand guarantees around hardware access, data separation and governance instead of blocking the deal outright.

Still, Nvidia has already learned that buying a trusted neutral layer can attract much more scrutiny than buying a normal software company.

Q13So why is Nvidia buying Hugging Face?

Nvidia is pursuing Hugging Face because controlling open-model distribution could protect its GPU business long after today's hardware lead starts facing stronger alternatives.

Put the pieces together and the motive is pretty clear. Nvidia's biggest customers are spending heavily on non-Nvidia accelerators, while three direct customers still represented 44% of Nvidia's first-half revenue. At the same time, Hugging Face is adding open models, datasets and applications quickly, converting more users into paying customers and sitting directly in the workflow where developers decide what to run.

Ownership would give Nvidia several ways to turn that position into an advantage. Hugging Face can make Nvidia deployment the easiest route for popular models, feed workloads into Nvidia-backed cloud capacity, push Nvidia's software deeper into open-model development and strengthen Nvidia's position in robotics and local AI before those markets settle around mature standards.

Paying roughly 86 times Hugging Face's annualized revenue looks extreme for a standalone software company, but Nvidia can fund the acquisition from a fraction of one quarter's profit. The bigger risk is damaging the developer neutrality that created Hugging Face's network in the first place.

Nvidia wants Hugging Face because the next fight in AI will be partly about who controls the route from models to compute. Hugging Face sits unusually close to that route today. Nvidia would rather own that crossroads than watch a cloud provider, competing chip company or another strategic buyer take it first.