Nvidia Stock ($NVDA): Hugging Face Strategy and What Actually Shipped
- Hugging Face’s price list is denominated in Nvidia part numbers. Seven of them — T4, L4, L40S, A100, H100, H200, B200 — are the GPU options, priced by the hour. That integration is quieter than any announcement and it is the one that stuck.
- Nvidia has become a large publisher on the Hub in its own right: 920 models, 54.2 million downloads. But 196 of those models, and half of those downloads, are other organisations’ models that Nvidia re-issued.
- Its single most-downloaded model is Alibaba’s. nvidia/Qwen3.6-35B-A3B-NVFP4 has 11.7 million downloads — a quantisation of Qwen, in NVFP4, Nvidia’s own four-bit format.
- The co-branded services from the partnership announcements — NIM serverless, Train on DGX Cloud, Training Cluster as a Service — are not on either company’s current product surface. Nvidia’s own DGX Cloud page does not mention Hugging Face at all.
There is a version of the Nvidia and Hugging Face story that is a list of joint announcements. This is not that, because most of those announcements are no longer findable on either company’s product pages.
What is findable is stranger and more interesting: Nvidia is now one of the biggest publishers on Hugging Face, and half of what people download from it is somebody else’s model.
What is the Nvidia–Hugging Face relationship actually made of?
A price list and a publishing account. Neither was the headline of any partnership announcement.
Hugging Face rents GPU time for Spaces and inference endpoints, and every option is an Nvidia part, priced by the hour: T4, L4, L40S, A100, H100, H200 and B200. The word “NVIDIA” appears 37 times on that pricing page. The word “DGX” — the brand under which the two companies announced their joint services — appears zero times.
The second half is the publishing account, and it is substantial.
Why is Nvidia publishing other companies’ models?
Because the format is the product. The top item on that chart is a quantisation: Nvidia takes an open model — here Alibaba’s Qwen — converts it to NVFP4, its own four-bit numeric format, using its own Model Optimizer toolchain, and republishes it.
NVFP4 runs best on Nvidia silicon. So the most convenient way to run the most popular open models at low precision arrives pre-packaged in a format tuned for the hardware Nvidia sells. Nobody has to be persuaded of anything; the convenient default simply routes through Nvidia’s software.
This is not hidden. Every one of those models declares its origin in a base_model tag, which is how they are counted here.
How big is this, in numbers you can check?
920 models, 54.2 million downloads, and half of that traffic is re-issued from elsewhere.
The asymmetry is the point. A fifth of the catalogue by count, half of it by use. And the largest single contributor is Qwen: 74 models, 16.2 million downloads — more, on its own, than every Google-derived, Meta-derived and DeepSeek-derived model in the account combined.
Which announced services can you use today?
Two of the five things on this list are on a live product surface. Three are not.
The NIM absence is the most checkable. Hugging Face documents its inference providers one page each: nebius, together, groq and cerebras all resolve. The equivalent addresses for nvidia and nvidia-nim return 404 — the same response as a provider name invented for the test. Ask the Hub’s API which providers serve that Qwen model and it names deepinfra, featherless-ai and scaleway. Nvidia is not among them.
Nvidia’s own DGX Cloud page, half a megabyte of it, does not contain the words “Hugging Face”.
None of that proves anything was cancelled. Services move behind enterprise sales, get renamed, or live somewhere we did not look. What it does mean is that a reader who goes looking today, on the pages both companies actually maintain, will not find them.
So is the partnership over?
No — it stopped being a partnership and became infrastructure. That is a demotion in press-release terms and a promotion in every other sense.
A co-branded service is a thing you can cancel. A price list denominated in your part numbers, and a model catalogue where the most convenient version of everyone else’s work arrives in your format, is not really cancellable by anyone. It is what people do by default.
Does any of this tell you what the stock is worth?
No, and this page will not pretend otherwise.
You can read the numbers above as evidence about a distribution position: how open-model developers actually get their models, and whose software sits in that path. That is a real thing to know and it is measurable, which is why it is here.
What it is not is a valuation. This site does not publish price targets, does not forecast where a share price goes, and does not republish other people’s targets even with a name and a date attached. None of what is measured above has a disclosed revenue line attached to it, so nobody outside the company can size it from public data — including us.
If you came here for a number to act on, the honest answer is that this page is upstream of that decision, not a substitute for it.
How can I check these numbers myself?
All of it is public and none of it needs an account.
- The catalogue: the Hub’s API answers
https://huggingface.co/api/models?author=nvidia&sort=downloads&direction=-1&limit=100&full=true, paginated through theLinkheader. Each model’s declared parent is in itstagsarray asbase_model:owner/name. - A single model:
https://huggingface.co/api/models/nvidia/Qwen3.6-35B-A3B-NVFP4returnscardData.base_modeldirectly. - Inference providers:
huggingface.co/docs/inference-providers/providers/<name>returns 200 for a real provider and 404 otherwise. - The price list: the GPU rows on Hugging Face’s pricing page.
One trap, since it cost us a wrong answer before it was caught: the list endpoint does not return cardData, even with full=true. Read the base_model: tag instead. Reading cardData.base_model from a list gives you None for every row and a very confident, very wrong zero.
How we verified this
🔴 This page contains no price target, no valuation and no forecast, and that is a rule rather than an omission. Everything measured here is a product figure: models published, downloads, whether a service appears on a live page. What any of it is worth to a shareholder is not a question this site answers.
🔴 Model classification comes from each model’s own declaration, not from its name. The first pass guessed by checking whether a model’s name contained “qwen” or “gemma”. That is a heuristic and it can be wrong in both directions. The figures here read the base_model: tag the publisher records on the Hub.
⚠️ That fix was forced by a silent failure worth recording. The second attempt read cardData.base_model from the Hub’s list endpoint with full=true. It returned nothing for all 920 models and produced the clean, confident, entirely false result that 0% of Nvidia’s catalogue derives from anyone else. The list endpoint carries no cardData at all; the declaration is in tags, as base_model:Qwen/Qwen3.6-35B-A3B. It was caught only because a single-model lookup had already shown otherwise.
✅ The absence of the NIM inference route was tested, not assumed. Provider pages for nebius, together, groq and cerebras all return 200. The equivalent paths for nvidia and nvidia-nim return 404 — the same as an invented provider name used as a control. Separately, the Hub’s API lists deepinfra, featherless-ai and scaleway as serving that Qwen model, and no Nvidia entry.
⚠️ “Not found today” is not “cancelled”. These services may exist under other names, behind enterprise agreements, or on pages we did not find. The claim made here is narrow and literal: they are absent from the current public product surfaces that were checked, which are named so anyone can check the same ones.
✅ The pricing-page finding was read rather than inferred. “NVIDIA” appears 37 times on Hugging Face’s pricing page and “DGX” appears zero times. The 37 are GPU instance rows — T4, L4, L40S, A100, H100, H200, B200 — with hourly rates beside them.
⚠️ Download counts are a snapshot, taken 28 August 2026. They move daily. The shape of the finding — a large minority of the catalogue, a majority of the traffic — is more durable than any individual number, and the last section shows how to re-run the count.