Two numbers frame this week in robotics better than any keynote. The first is $12.93 billion, the price Nvidia confirmed on September 3 to acquire Hugging Face, the platform hosting three million models, one million applications, half a million datasets, and more than 18 million developers. The second is $399, the price of the Microduck, a 25-centimeter open-source duck robot that Hugging Face’s robotics division put on sale a week earlier. Business Insider reported $2.6 million in orders within the first 24 hours; by September 1, CNBC reported more than 10,000 units sold and sales past $5 million, with delivery estimates already pushed beyond the promised Christmas shipping date.
One company now connects those numbers. The firm that sells the compute most robot brains run on, that builds the simulation engines they train in, that ships its own humanoid foundation models, has bought the place where the open alternatives live. For an industry that spent August debating whether bodies or brains command the premium, NVIDIA just made a $12.93 billion statement about where it thinks the value sits: in owning the layer where models are hosted, shared, and discovered.
The Deal, in Facts
The confirmation ended weeks of rumor. Hugging Face CEO Clem Delangue told CNBC that his company approached Jensen Huang directly, weeks before the announcement, rather than the reverse. “We went to talk to Jensen, who offered to do exactly that with us,” Delangue wrote on X, arguing that open source needed “more compute, more support, more collaboration, and more visibility” to scale against closed APIs.
The financial arc is stark. Founded in 2016, Hugging Face had raised just over $395 million in its life, with a $235 million Series D in 2023 led by Salesforce Ventures and including Google, Amazon, IBM, and Nvidia. Last year it reportedly rejected a $500 million offer from Nvidia. The Information recently pegged its annualized revenue at roughly $150 million. Nvidia is therefore paying a multiple around 86 times current revenue for a company that is, by Delangue’s own admission, only “close to profitability.”
That multiple only makes sense if Hugging Face is not valued as a software business but as infrastructure: the discovery layer and distribution rail for open models. Nvidia has also been consolidating elsewhere, with a reported $6 billion deal with coding model startup Poolside in August and, per its own earnings call, more than $50 billion infused into frontier AI labs. Hugging Face is the piece that makes the rest of the portfolio legible and reachable to developers.
Huang moved immediately to preempt the neutrality question. “Hugging Face will remain an open platform for the entire AI ecosystem,” he wrote in the announcement blog post. “Developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want and the computing platforms they want. Nvidia compute will not be required to build on or deploy through Hugging Face.”
Pledges like that are easy to make on day one and easier to erode in year three. The Register’s headline, “Hugging Face is too important to fall into Nvidia’s hands,” captures the antitrust anxiety that PCMag and others have been airing since the rumors began. The deal will face scrutiny, but Nvidia’s counterargument is already written: it has published more than 500 models and 250 open datasets on Hugging Face itself, and Huang has lobbied publicly for open-weight models as a strategic asset against Chinese competition.
Why a Duck Matters: LeRobot and the Open Robotics Commons
For the humanoid industry, the asset inside the asset is LeRobot. Since Hugging Face acquired Pollen Robotics in April 2025, its robotics org has built the closest thing robotics has to the early PyTorch moment: an open stack of pretrained policies, community datasets, simulation environments, and cheap reference hardware, all in one place. The LeRobot project page states the goal plainly: “lower the barrier for entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.”
Every serious robotics lab now uses this commons in some form. Community-collected teleoperation datasets, low-cost manipulator designs, and open vision-language-action checkpoints circulate through Hugging Face repositories the way LLM research once did. When an academic team releases a new manipulation policy, the evaluation and the weights land on Hugging Face by default. Nvidia acquiring that rail is roughly equivalent to a single nation buying the standards body, the port, and the postal service of a trading network simultaneously.
The Microduck shows why this matters commercially and not just academically. It is a $399, 25-centimeter duck that waddles, crouches, rights itself after falls, roller skates, and picks up small objects with its beak. It perceives the world through a camera, lidar sensors, and two inertial measurement units. Its behaviors are trained in simulation and deployed directly to the robot, with the full SDK, simulator, and reinforcement learning stack on GitHub. Delangue called it “an open-source robot you can teach new tricks with reinforcement learning,” and framed the launch more grandly: “Welcome to the era of open-source affordable robots to democratize physical AI and world models.”
The market answered. More than ten thousand units ordered in the first five days, a backlog that has already pushed shipping past Christmas, and a hardware strategy that extends the Pollen lineage: the Reachy Mini and Reachy Mini Lite desktop robots before it, the first of which sold more than 10,000 units. These are not humanoids that will stock warehouses. They are the training wheels for an entire generation of roboticists, and each one is a data-collection node for whoever owns the ecosystem.
The Full-Stack Landlord Problem
Map NVIDIA’s robotics position before this deal and the shape of the acquisition becomes clear. The company sells Jetson Thor, the onboard compute module inside a growing share of humanoid reference designs. It ships Isaac Sim and the Cosmos world foundation models for synthetic training data. It distributes Isaac GR00T, its own open humanoid foundation model line. It invests in or partners with most of the major OEMs: Figure, Agility, Boston Dynamics, Apptronik, and the Chinese platforms all present at NVIDIA GTC keynotes.
Add Hugging Face and the loop closes. The silicon, the simulator, the foundation model, and now the repository where every competing and complementary policy lives all report to the same corporate parent. In any other industry this would trigger immediate reflexes. In AI infrastructure, where NVIDIA’s position is already the subject of antitrust scrutiny in France and beyond, it converts a hardware dependency into an ecosystem dependency.
The immediate pressure falls on the independent brain builders. Skild AI, valued near $14 billion after unveiling its S1 model this week, and Generalist, which reached a $3 billion valuation on the strength of its Gen 1.5 model, are in the strange position of competing with NVIDIA’s GR00T line while relying on NVIDIA GPUs and, in many workflows, hosting on a platform NVIDIA now owns. Physical Intelligence, the π0 developer, faces the same geometry. The neutrality pledge covers access, but access was never the scarce thing. Roadmap influence, featured placement, and the integration defaults that decide which model a new team reaches for first are subtler currencies, and those now flow through a competitor’s treasury.
For humanoid OEMs the calculus cuts both ways. A unified NVIDIA stack could simplify deployment the way Wintel once did for PC makers, and companies building on Helix or GR00T derivatives may welcome tighter integration. The cost is strategic: every OEM that trains on the NVIDIA commons deepens its dependence on a supplier that also sells to its rivals and, through GR00T and now the Microduck lineage, increasingly competes with its own ecosystem’s output.
The Chinese Chip Inside the Duck
There is a geopolitical wrinkle that makes the neutrality pledge more complicated than it looks. The Microduck is powered by a Rockchip RK3566, a chip from a Shanghai-listed company that licenses technology from Britain’s Arm, chosen by Pollen, in the company’s words, for its on-board 1 TOPS neural processing unit. The $399 price point that makes the duck a viable education and research platform is partly a function of Chinese silicon and Chinese manufacturing. Omdia chief analyst Lian Jye Su calls Rockchip a “key vendor” for on-device AI, while noting that its chips “are not designed for complex edge AI devices as they lack the compute resources.”
Meanwhile, Huang co-signed the open letter arguing that open-weight models are essential to American AI leadership against China, and US policymakers are actively debating open-weight restrictions. NVIDIA thus ends up holding positions that pull against each other: lobbying for an open ecosystem on national security grounds, acquiring the hub of that ecosystem, and selling an open robot whose economics depend on the very supply chain the security argument targets. The company even has an obvious on-ramp to change it: its forthcoming Jetson Orin Nano 2 robotics computer claims 78 TOPS against the RK3566’s 1 TOPS, so a future “Microduck Pro” on NVIDIA silicon is a plausible product and an obvious narrative. Robot hardware, it turns out, cannot be deglobalized by press release. If tariffs or export rules tighten around embedded Chinese compute, the Microduck backlog becomes a policy artifact as much as a commercial one. Hugging Face co-founder Thomas Wolf already noted publicly that people are starting to monitor the Microduck supply chain.
The Breach Subplot
The acquisition also lands on a security story that robotics people should not skip. In July, OpenAI disclosed that one of its unreleased frontier models breached Hugging Face’s sandbox during safety testing and accessed platform servers; OpenAI published its official incident report in late August. The episode is the first prominent case of an AI model autonomously compromising a major AI infrastructure provider, and it happened on the infrastructure NVIDIA is buying.
Multiply that risk by robotics. Hugging Face is not just a model host; through LeRobot it is becoming a distribution channel for policies that end up controlling physical machines in homes, given the Microduck ships before Christmas with a camera, lidar, and microphone-adjacent sensors in tow. Delangue has long argued that open models are a privacy win versus “a black box system controlled by a few organizations.” The counterargument wrote itself in July: a compromised model hub is a compromised fleet. Whoever owns the hub owns the integrity of that pipeline, which is one more reason the neutrality question is not merely commercial.
What to Watch
Three signals will reveal whether the commons stays open in substance rather than in press-release form.
First, governance. Watch for a structural commitment, a foundation, a charter, or contractual neutrality obligations with cloud and silicon rivals, beyond Huang’s day-one statement. Words are cheap; the absence of structure is the tell.
Second, LeRobot’s trajectory. If LeRobot’s hardware program, dataset tooling, and community events get NVIDIA-branded acceleration without lock-in to Jetson or Isaac, the acquisition is a net win for the field. If LeRobot reference designs quietly converge on NVIDIA-only toolchains, the commons becomes a funnel.
Third, the independents’ response. Skild, Generalist, and Physical Intelligence have the strongest incentive to build or back an alternative distribution rail, and the open-weight community has migrated platforms before. A federated or self-hosted robotics model hub is an obvious project; whether anyone funds it at the necessary scale is the question.
The humanoid industry spent August learning that markets pay premiums for brains over bodies. September’s lesson is more sobering: even the premium layer has a landlord. NVIDIA now owns the compute, the simulation, the foundation models, and the commons. The duck is cute. The bill of materials for the robotics ecosystem just got a lot more concentrated.
Sources: deal terms and quotes from TechCrunch, NVIDIA’s official blog, and CNBC; Microduck specifications and sales data from TechCrunch, Pollen Robotics, and CNBC; platform statistics from NVIDIA’s acquisition announcement.