Nvidia has signed a definitive agreement to buy Hugging Face for about $17.9 billion, and filed it with the United States Securities and Exchange Commission on 2 September. Hugging Face is the repository where developers publish and download open weight AI models, and it sits underneath a large share of the local AI work being done in Australian businesses right now.
This piece first ran on 29 August, when the deal was a report and neither company had commented. It is now signed. What is left is regulatory approval, with completion expected in the first half of 2027.
What the filing actually says
The filing puts the total transaction value at about $17.9 billion. Roughly $16.5 billion of that is the purchase price payable to Hugging Face shareholders, subject to adjustment, with about $1.4 billion more in a retention stock programme for the people who stay, according to the filing. The original report from The Information on 26 August had the number right.
Nvidia says Hugging Face will remain an open platform for the entire AI ecosystem, and that Nvidia compute will not be required to build or deploy models on it. The platform it is buying carries more than 18 million developers, 3 million models, 1 million applications and 500,000 datasets.
It is Nvidia’s largest acquisition by a wide margin. The previous record was Mellanox, the Israeli networking company Nvidia bought in 2020 for about $9.6 billion, a deal that went on to underpin its data centre business. Hugging Face was founded in 2016 and was last valued at roughly $6.3 billion in a 2023 round that Nvidia itself took part in.
It would also be the second large consolidation of AI plumbing this month. Stripe bought OpenRouter, the service that routes API calls across model providers, for more than $9.7 billion earlier in August. The layer that sits between developers and models is being bought up quickly, and the buyers are not the model makers.

Why NVIDIA wants the model hub
The strategic logic is defensive. OpenAI, Google, Amazon and Anthropic are all building their own silicon, which chips away at the assumption that serious AI work has to run on Nvidia hardware. Open weight models cut the other way, because every team that downloads one and runs it locally still needs GPUs to do it.
Nvidia has circled the company before. It offered about $695 million in late 2025 at a valuation near $9.7 billion and was turned down, according to CNBC. Hugging Face has since grown revenue to roughly $208 million a year and its chief executive has said it is approaching profitability, which explains the gap between the two numbers.
Why Australian businesses should care
For Australian companies the interest is not the price. It is that Hugging Face is the default supply route for open weight models, and open weight models are the practical answer for teams that cannot send customer data to an overseas API. That is the same reasoning behind running Alibaba’s Qwen on your own servers.
A repository owned by a chip vendor is not the same as a neutral one, even if nothing changes on day one. Licensing terms, hosting costs and what gets promoted on the front page are all decisions an owner makes. Any business that has built a data residency position on self hosted models should know who controls that supply, which is the practical end of the CLOUD Act and data residency question.
Worth watching, not worth reacting to
Nothing needs to change this week. The weights already downloaded keep working, the open licences on them do not retroactively change, and no announcement has been made. The sensible move is to know which models your systems depend on and where else those weights are mirrored.
That inventory is cheap to build now and expensive to build in a hurry, and it is the sort of question worth settling before it becomes urgent rather than after. It sits naturally alongside the rest of an AI strategy review.
