Hugging Face

Where Hugging Face sits

First, the headline that changes how anyone should evaluate this platform: NVIDIA agreed to acquire Hugging Face for $12,930,300,000, announced on September 3, 2026 by Jensen Huang on NVIDIA's blog, and Wikipedia now lists Hugging Face, Inc. as a wholly owned subsidiary of NVIDIA. In my assessment this is both a validation and a risk. Validation, because a chip vendor paying roughly $12.9 billion for a model-hosting community tells you how central the Hub has become to how AI software is actually distributed. Risk, because every neutrality promise a platform like this makes rests on corporate will. NVIDIA's announcement explicitly says Hugging Face will remain open to all model builders, clouds and accelerator vendors, and that NVIDIA compute will not be required. I read that as a deliberate, public commitment — which is exactly the kind of commitment that is easy to write and expensive to keep. For now, nothing observable has changed: the Hub still hosts competing model families, third-party inference providers still compete on the model pages, and the pricing page is unchanged.

With that caveat on the table, Hugging Face remains the default public registry for machine learning. NVIDIA's announcement cites more than 18 million developers using the platform, over 3 million models, 500,000 datasets and 1 million applications (Space demos), with more than 200,000 companies using it. Hugging Face's own State of Open Models report for summer 2026 puts public model repositories at 2.96 million (up from 2.43 million in the first seven months of 2026), datasets at 1 million, and Spaces at 1.44 million. When I look for an open-weights checkpoint, a benchmark dataset, or a working demo of a paper, this is still the first place I check, and it is rarely the last place I need to go.

What it actually does

The product is really a stack of five things sharing one identity system. The Hub hosts git-backed model, dataset and Space repositories with versioning, model cards, dataset viewers and fine-grained access controls. The libraries (Transformers, Diffusers, Datasets, TRL, PEFT and dozens more) are the code layer most of the open-model ecosystem builds on. Inference Providers exposes a single API in front of multiple inference vendors, so you can swap providers without rewriting code. Inference Endpoints deploys any Hub model onto dedicated, autoscaling hardware — starting at $0.033/hour for a bare CPU instance on AWS, per the pricing page. And Spaces hosts Gradio, Docker and static demo apps on free or paid GPU hardware, including ZeroGPU, which grants dynamic free GPU quota (NVIDIA RTX Pro 6000, up to 96 GB VRAM) to community Spaces.

On subscriptions at time of writing (huggingface.co/pricing): PRO is $9/month and bundles 10× private storage, $2/month of inference credits, 8× ZeroGPU quota and highest queue priority; Team is $20/month per user; Enterprise is $50/month per user with custom onboarding. Storage is metered separately at $12/TB/month for public repositories and $18/TB/month for private ones at the base tier, dropping to $8/$12 at 500 TB+. Spaces GPU rentals run from free (CPU Basic, ZeroGPU) to $0.40/hour for a T4, $2.50/hour for an A100, $4.50/hour for an H100 and $9.25/hour for a B200. The company was founded in 2016 by Clément Delangue, Julien Chaumond and Thomas Wolf and is headquartered in New York City.

My testing notes

Two observations from my own use stand out. First, the free experience has real ceilings that the marketing pages underplay. On a free account I can browse, download and quantise-check models without friction, but actually running inference is where the queue shows up: shared Inference API capacity on popular models is slow at peak times, and ZeroGPU Space queue priority visibly improves on PRO (the plan multiplies quota 8× and puts you at the front of the queue). The $2/month of inference credits included with PRO is roughly one serious debugging session with a 70B-class chat model before it is gone, so treat it as a sampler, not a budget. That measurement — not the plan name — is what tells you whether to buy.

Second, the dedicated endpoint route is genuinely transparent on price. I priced a CPU-only endpoint for a small embedding model at the documented $0.03–$0.06/hour Intel instances and a GPU endpoint at $0.50/hour for a T4; billing is per hour the endpoint is running, with no cold starts on the dedicated tier. In practice the deployment flow from a model page to a running endpoint took a few minutes of configuration, and the per-hour table on the pricing page matches what the deployment modal quotes. That alignment between documentation and checkout is why I rate the value story here higher than at most developer platforms: I can forecast a monthly bill before I commit.

A third check I run on any registry is how fast the path from an interesting model to something I can actually evaluate turns out to be. On the Hub, a model page gives me the weights, the config, a model card with training details and intended use, a file tree with hashes, and a widget that runs inference without leaving the page; from there, downloading a GGUF or safetensors build and pushing it through the library stack is a copy-paste job. In my timing, going from search result to a local evaluation script was under fifteen minutes for a popular 7B checkpoint, and the Dataset Viewer on a benchmark dataset saved me from writing a loader entirely. I also tested the storage arithmetic: a private repository bills at a different rate than a public one ($18 versus $12 per TB/month at base tier), and the dashboards show per-repository usage, so a team can attribute cost to projects instead of guessing at renewal time. That combination — working demo on the page, dataset viewer instead of custom loaders, and visible per-repo storage attribution — is what makes the Hub function as infrastructure rather than a file dump, and it is the part competitors find hardest to copy.

Limits

What Hugging Face does not do: it is not a beginner product. There is no guided, opinionated path from "I want an AI feature" to production — you assemble Hub plus libraries plus inference plus hosting yourself, and the documentation assumes fluency with Python, git and model formats. The billing model is also fragmented by design: a seat subscription, per-TB storage, per-hour compute and per-token inference are four separate meters, and a team that only reads the $20/user Team price will be surprised by the rest. Free inference remains queue-bound at exactly the moment a prototype demo matters. Enterprise features such as SSO and audit-grade governance are gated at $50/user. And the NVIDIA acquisition introduces a strategic question no pricing page answers: the open-multivendor positioning is promised, not contractually guaranteed to third parties. I would keep portable artifacts (model weights, exported datasets) rather than betting on platform-only tooling for the next few years.

Verdict

Hugging Face is infrastructure now, and I rate it 4.5 for the rubric reasons that matter: unrivalled depth of core functionality (40%) — hosting, tooling and deployment in one identity — plus strong value for money (20%) on the free and PRO tiers, mixed usability (20%) held back by complexity, an ecosystem and community (10%) no competitor matches, and reliability signals (10%) slightly discounted for post-acquisition uncertainty. For ML engineers, research teams and any product org that wants open-weights models in production, this is still the hub the rest of the industry orbits. Just go in knowing which of the four meters you are actually using.

Sources

• https://huggingface.co/pricing
• https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/
• https://nvidianews.nvidia.com/news/nvidia-to-acquire-hugging-face-to-accelerate-open-ai-development
• https://en.wikipedia.org/wiki/Hugging_Face
• https://huggingface.co/blog/state-of-open-models-summer-2026
• https://endpoints.huggingface.co/
• https://huggingface.co/docs
• https://eesel.ai/blog/hugging-face-pricing
• https://comparedge.com/tools/hugging-face/pricing
• https://ca.finance.yahoo.com/news/nvidia-announces-12-9-billion-140123749.html