Nvidia has agreed to acquire Hugging Face for $12.9 billion, valuing the ten-year-old open-source AI platform at roughly 86 times its $150 million in annualized revenue. The deal, reported on August 27, 2026, gives Nvidia control of a hub hosting more than 3 million open-weight large language models and roughly 4 million total models, along with a growing cloud, storage, and robotics business that had not existed in this form a year ago.<\/p>\n
The acquisition follows a January report that Hugging Face had turned down a $500 million investment from Nvidia at a $7 billion valuation. At the time, independence appeared to matter more than capital. Nvidia returned with a full takeover offer, nearly tripling the $4.5 billion private valuation Hugging Face carried in 2023.<\/p>\n
The price is small relative to Nvidia’s scale. The chipmaker reported $96.2 billion in quarterly revenue on the same day the deal leaked and forecast a 70 percent jump in revenue next fiscal year. It has also disclosed $18 billion committed to equity investments through 2027. At that pace, $12.9 billion equals roughly thirteen days of sales, a rounding error in exchange for ownership of the distribution layer for open-weight AI.<\/p>\n
Nvidia’s strategic interest goes beyond models. Hugging Face already operates a managed compute layer on top of its hub. Developers can spin up Inference Endpoints, ranging from $0.03-per-hour CPU instances to $80-per-hour clusters of eight Nvidia H100 GPUs. They host interactive demos on Spaces, billed by the hour for GPU hardware, and route API calls through Inference Providers, a billing layer connecting users to third-party GPU clouds such as Together AI, SambaNova, and Groq.<\/p>\n
The most important piece is Training Cluster as a Service, a joint product between Hugging Face and Nvidia that gives any of the platform’s 500,000 organizations on-demand access to large GPU clusters, billed only for the duration of a training run. It functions as a distribution channel for Nvidia compute packaged as a developer tool.<\/p>\n
Hugging Face is now a vertically integrated AI platform spanning frontier models, physical robotics, and enterprise storage, built on a headcount of 250 people. The company has made six acquisitions, all small and talent-driven.<\/p>\n
The robotics business started as a software problem. After releasing an open-source library for robotics, the team hit a hardware bottleneck. As co-founder Thomas Wolf put it: “Hardware was very expensive. Even the cheapest ones are still like several thousand dollars, $7,000 to $10,000, $20,000 to $30,000, $50,000.” The first affordable entry point was a $100 robotics arm. It sold more than 10,000 units, and the company is on track to sell 20,000 this year.<\/p>\n
Hugging Face followed that by acquiring French robotics startup Pollen Robotics, which now forms the core of its physical AI division. The second consumer kit, the Reachy Mini, is priced between $399 and $499 and is set for release this summer. On the same day the Nvidia deal was reported, Wolf unveiled Microduck, a 25-centimeter open-source biped with 15 actuators and a full sensor suite: camera, speaker, LiDAR, NFC, Bluetooth, and Wi-Fi. Designed for training from scratch with reinforcement learning, it ships with more than half a dozen pre-trained policies so it can walk, sit, crouch, roller-skate, pick up objects with an articulated beak, and recover on its own, all for $399. Microduck’s order volume reached more than $2.6 million after the launch announcement.<\/p>\n
Enterprise storage has emerged as a quiet growth driver. Because the hub houses petabytes of model weights and datasets, the engineering team built an ultra-efficient blob storage system. Wolf said the company is now selling it to customers storing petabytes or terabytes of data, and the traction is strong. In June, AI lab Arcee became the first major American company to replace AWS S3 with Hugging Face Private Storage in a multi-million dollar commercial partnership.<\/p>\n
Wolf sees open-weight models closing the gap with frontier closed systems faster than most observers expected. He points to GLM 5.2 by Chinese AI lab Z.ai, released weeks before, as the latest shock, calling it “surprisingly close to Opus 4.8 or Frontier.” His personal favorite is Google’s Gemma 4, which he describes as “really good” and small enough to run locally.<\/p>\n
He has also been experimenting with agent collaboration at scale. Wolf spent a week with more than 100 AI agents working freely on an open project. The result was a 5x inference speedup on Gemma 4 inside vLLM, what he called one of the most interesting emergent behaviors he has seen from agent swarms.<\/p>\n
Wolf’s path from theoretical physics to patent law to quantum computing shapes where he points the company. He sees AI driving progress across biology, materials discovery, mathematics, and physics, and is backing initiatives in each. Asked which field will be disrupted most, his answer was immediate: “All of them. I think research will change a lot.”<\/p>\n
He is also weighing the consciousness question raised by Anthropic’s recent paper on verbalizable representations forming a global workspace in language models, musing that “maybe these AI models have a consciousness of their own, or like a workflow where they find a workspace in their mind where there is a concept.”<\/p>\n
On automation anxiety, Wolf’s answer is direct, then hedged: “That’s the goal, to stop going to work. I don’t know, but not in the super short term.” He expects humans with taste and context-setting skills to remain essential, and thinks the fully automated enterprise will take longer than current hype suggests.<\/p>\n
Nvidia bought the distribution layer for the one corner of AI it did not control: open-weight development. The cloud, storage, and inference businesses inside Hugging Face are still small in Nvidia terms. Cloud services, storage, and subscriptions produced roughly $150 million in annualized revenue as of this summer, up from about $100 million two months earlier.<\/p>\n
But Nvidia’s own DGX Cloud offering was reportedly scaled back about a year ago, and the chipmaker has promised to help cover tens of billions of dollars in cloud computing deals for its largest customers. If those customers do not use all the compute they signed up for, Nvidia could end up with excess capacity. Owning Hugging Face gives the company a built-in customer base of millions of developers and thousands of enterprises, already renting GPUs, already paying for inference, already trusting the brand, ready to absorb that spare capacity.<\/p>\n
As of August 29, 2026, the deal has not yet closed and remains subject to regulatory approval. Financial figures and product details reflect the most recent publicly available information.<\/p>\n
Nvidia has agreed to acquire Hugging Face for $12.9 billion, valuing the company at roughly 86 times its $150 million in annualized revenue.<\/p>\n
Hugging Face hosts an open-weight model hub with more than 3 million large language models and roughly 4 million total models, runs Inference Endpoints, Spaces, and Inference Providers for GPU-backed AI workloads, sells enterprise storage through Hugging Face Private Storage, and ships affordable open-source robotics kits including the Reachy Mini and Microduck.<\/p>\n
Hugging Face gives Nvidia a built-in distribution channel for its GPUs, a base of millions of developers and 500,000 organizations already renting compute through Training Cluster as a Service, and a way to absorb excess GPU capacity if large cloud commitments go underused.<\/p>\n