Based on Hugging Face and NVIDIA announcements and reporting by The Information and Young Research. OpenAI’s in-house humanoid effort and the reported NVIDIA-Hugging Face deal are unconfirmed reports, described as such below.
After the model layer commoditized, the contest is moving to the body โ and the same open/frontier barbell that organized AI software is reassembling, one layer down, in atoms.
How The Robotics Layer Is Forming โ Key Markers
Dec 2024
The Information first reports OpenAI discussed building its own humanoid robot โ an internal direction, not a product, and still unshipped as of this writing.
Early 2025
Figure ends its collaboration with OpenAI and ships its own Helix vision-language-action model. The assumed “OpenAI brain, startup body” division of labor formally collapses.
~July 2026
Hugging Face and NVIDIA announce a partnership integrating Isaac GR00T (open VLA), Teleop, Jetson Thor, and DGX Spark into the LeRobot library โ making NVIDIA the substrate under the open robotics commons.
Aug 2026 โ this week
Hugging Face launches Microduck, a $399 educational AI robot โ its second product after Reachy Mini. CEO Clement Delangue targets ~50,000 units sold. Separately, The Information reports a ~$12.9B NVIDIA acquisition of Hugging Face: agreed but unsigned, unconfirmed, and subject to regulatory review.
Hugging Face, NVIDIA, The Information, and Young Research are the sources this piece draws on. Hugging Face this week launched Microduck โ a $399 AI-powered educational robot, its second hardware product after the Reachy Mini โ with CEO Clement Delangue publicly targeting around 50,000 sales. That is a goal, not a result, and Microduck is an accessible educational device, not a frontier humanoid. But the launch is a useful peg, because it marks where Hugging Face has decided to plant its flag: cheap, open, developer-first, and built on NVIDIA’s physical-AI stack.
The NVIDIA substrate is not incidental. Through the LeRobot partnership announced around July 2026, NVIDIA’s Isaac GR00T open vision-language-action model and its Teleop framework were integrated directly into LeRobot, Hugging Face’s open robotics library. The Reachy 2 humanoid is interoperable with NVIDIA’s Jetson Thor computer; the Reachy Mini (and by extension, the Microduck tier) runs on DGX Spark, with NVIDIA’s Cosmos physical-AI model planned behind them. NVIDIA is not a partner at the margin โ it is the compute and model layer underneath Hugging Face’s entire open-robotics stack.
On the frontier side, the picture is different โ and notably, it does not involve OpenAI’s models running inside anyone else’s robot. The Information has reported, first in December 2024 and in subsequent coverage, that OpenAI discussed building its own humanoid and has reportedly stood up an internal robotics division, hiring hardware and manufacturing engineers toward that end. That effort is described as early and unshipped; treat it as a reported direction, not a product announcement. Figure, which once collaborated with OpenAI, ended that relationship in early 2025 and now runs its own Helix VLA system. 1X runs its own Redwood AI. Tesla has Optimus. The frontier has gone vertical โ building both body and brain in-house โ which means OpenAI’s current influence on shipping humanoids is as a would-be builder, not as the intelligence layer inside everyone else’s robot.
The key insight: The open/frontier split that organized the model layer โ an accessible commons on one side, closed vertically-integrated stacks on the other โ is now reassembling in hardware. NVIDIA sits underneath the open side exactly as it sits underneath the model commons. And the reported (unconfirmed, unsigned) NVIDIA acquisition of Hugging Face, at roughly $12.9 billion per The Information, would hand NVIDIA not just compute but the distribution hub sitting on top of it.
The Structural Read
The Map of AI framework describes the AI stack as layered: compute at the base, models in the middle, applications and distribution at the top. What Microduck’s launch clarifies is that embodiment is becoming a new contested layer โ one being structured by the same competitive logic that organized every layer above it.
The model barbell was: open-weight models (Llama, Qwen, Mistral) on one side, giving away capability to commoditize the layer; frontier closed models (GPT-4o, Claude, Gemini) on the other, racing to stay ahead of what open can replicate. The robotics barbell forming now maps onto that cleanly. Hugging Face is the open side: cheap robots, an open library (LeRobot), embodiment as a commons to widen rather than a moat to hoard. The frontier humanoids โ OpenAI’s reported in-house effort, Figure/Helix, 1X/Redwood, Tesla/Optimus โ are the closed side, and they have all made the same structural choice: vertical integration. Brain and body are one product. Make, not buy.
That make-vs-buy swing is worth dwelling on. A year ago, the conventional read was that AI labs would supply cognition and hardware startups would supply bodies. Figure’s decision to drop its OpenAI collaboration and build Helix internally was the clearest signal that the frontier rejected that model. You cannot optimize across the full stack if you don’t own it. The brain and the body have to be trained together, evaluated together, and iterated together โ and no partnership agreement makes that fast enough.
Map of AI โ Embodiment Layer
NVIDIA Is the Common Substrate โ Again
On the model layer, NVIDIA sells compute to both open-weight labs and frontier model companies โ it collects regardless of which side wins. On the robotics layer, the same geometry is forming. Hugging Face’s open robots run on NVIDIA’s GR00T, Jetson, and Cosmos stack. The frontier humanoids run their own models but on hardware that still leans heavily on NVIDIA silicon. If the reported NVIDIAโHugging Face acquisition closes โ and it is, as of this writing, agreed but unsigned, unconfirmed, and subject to regulatory review โ NVIDIA would own the open-robotics distribution hub sitting on top of the compute it already sells. That is not just a financial return; it is a structural lock on whichever tier of the robotics market scales first.
One hedge to hold clearly: the NVIDIAโHugging Face deal, reported by The Information at roughly $12.9 billion, is unconfirmed. It could collapse before signing. It would face regulatory scrutiny. The structural analysis of what that deal would mean is contingent on a transaction that has not happened. What is real and announced is the LeRobot partnership โ the compute and model integration is already in place; the ownership question is still open.
And a second hedge on scale: humanoid robotics is early and hard. The eye-catching market-size projections that accompany every robotics announcement are left out here deliberately โ they are not verifiable and they tend to obscure more than they reveal. What is observable is the structural choice each player is making, and that is the durable signal.
Three Implications
IMPLICATION 1 โ THE OPEN SIDE WIDENS THE MARKET, NOT THE MOAT
Hugging Face’s playbook with Microduck and LeRobot is the same one it ran with models: lower the cost of entry until the developer community does the distribution. A $399 educational robot is not a product that wins on margin โ it is a developer-acquisition strategy. The goal is to make LeRobot the default open robotics library the way the Hugging Face Hub became the default model repository. If that works, the open side generates a robotics commons, and NVIDIA (as the compute substrate beneath it) captures the economic upside without needing to win the robot market directly.
IMPLICATION 2 โ VERTICAL INTEGRATION IS THE FRONTIER’S ONLY VIABLE STRATEGY
Figure’s move away from OpenAI’s models toward its own Helix system was not a statement about OpenAI’s quality โ it was a statement about the optimization surface. A humanoid robot that acts in the physical world needs to be trained end-to-end: the visionmodel, the action policy, and the hardware dynamics have to be iterated together. Any seam in that stack โ any point where you hand off to a partner โ slows the feedback loop. That is why every serious frontier humanoid player (Figure, 1X, Tesla, and reportedly OpenAI itself) has converged on the same answer: own the full stack. The make-vs-buy question in robotics has a structural answer, and the frontier has found it.
IMPLICATION 3 โ OPENAI’S REPORTED ROBOTICS BET IS THE MOST UNCERTAIN VARIABLE
If OpenAI’s reported internal humanoid effort ships, it would enter a market where the competitive benchmark is already set by players who have been building in-house for years โ Figure, 1X, and Tesla have real robots in real environments. OpenAI would arrive late to the hardware side but potentially early to the AI cognition side, if its foundation models give it an advantage in the robot brain. That is a meaningful open question. But it is worth being precise: as of today, OpenAI’s robotics program is a reported direction and a hiring signal, not a product. The gap between a robotics division and a shipping humanoid is enormous, and the timeline is genuinely unknown.
Business Engineer Framework
The Map of AI โ Where Embodiment Fits in the Stack
The Map of AI tracks 9 layers and 200+ companies to identify where competitive leverage accumulates. The robotics layer is now forming, and the same structural forces โ open commons vs. closed frontier, compute as the durable substrate โ are visible one level down. The framework helps you read which players are building defensible positions and which are building bridges for someone else’s toll road. The NVIDIA-as-substrate pattern, and what the reported Hugging Face acquisition would mean for the full stack, is analyzed in depth in Beyond NVIDIA’s Moat.
A $399 educational robot does not move markets on its own โ but Microduck’s launch, read alongside the LeRobotโNVIDIA partnership, Figure’s in-house pivot, and the reported OpenAI robotics build, reveals that the structural contest in AI has shifted from producing tokens to acting in the world, and that the same organizing logic โ open commons below, vertically integrated frontier above, NVIDIA collecting underneath both โ is reassembling in hardware exactly as it did in software. The specific claims will firm up or fall away as OpenAI’s program develops and the NVIDIAโHugging Face deal either closes or collapses. The pattern they are early evidence of is already clear.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
