Robotics companies raised over $10 billion in 2025, yet the models powering their robots train on fewer than 5,000 hours of combined open-source real-world interaction data. Language models consume trillions of tokens scraped from the web. Physical AI has no equivalent. Every training example must be physically collected, one robot manipulation at a time.
That asymmetry is now the most expensive problem in AI.
The constraint is structural. Unlike text or images, robotic manipulation data cannot be crawled from the internet. It requires embodied hardware, human demonstrators, and annotators who understand task structure, failure modes, and semantic intent. Closing that gap is what makes data labeling for physical AI a distinct market from anything that came before it.
The Venture Thesis
Investors have noticed. Robotics funding hit $8.5 billion in 2025 through September alone. But the dollars are almost entirely stacked against foundation model developers, hardware manufacturers, and humanoid startups. The infrastructure layer that makes those models trainable, specifically, the physical world data supply chain, remains underfunded relative to the problem size.
Bessemer Venture Partners made this explicit in its April 2026 robotics outlook, where a former Waymo researcher wrote: the data problem in robotics is nowhere near solved. Closing the gap between 99% and 99.9% reliability is a steep hill that takes longer than most investors realize.
Scale AI grasped the opportunity early. The company launched its Physical AI Data Engine in September 2025, logging over 100,000 production hours at its San Francisco lab with clients including Physical Intelligence and Cobot. Meta’s $14.3 billion acquisition of a 49% stake in Scale at a $29 billion valuation in June 2025 made the data infrastructure bet explicit: whoever controls the ground truth for physical AI controls the training flywheel.
Market Map: Three Competing Approaches
Three distinct strategies are now competing to become the standard data stack for physical AI:
The real-world approach rests on a straightforward claim: robots learn dexterity from watching humans. Scale AI built collection infrastructure to capture those demonstrations at industrial volume, pairing them with semantic annotations encoding intent and failure modes. Physical Intelligence invested heavily in its own data flywheel, collecting proprietary interaction data across eight robot embodiments before releasing its pi-zero foundation model.
Emerging players are taking the approach further. Ground Truth Machine (groundtruthmachine.com) treats physiological signals as a calibration layer on top of behavioral demonstrations, capturing the gap between what a human demonstrator intends and what their body actually does. That signal, absent from every major existing dataset, is what the company calls the Authenticity Gap: the measurable divergence between explicit task instruction and implicit physiological ground truth. For training robots to handle edge cases in real human environments, that divergence may be the most informative data point in the stack.
NVIDIA’s synthetic bet is the largest in raw compute terms. Isaac Sim paired with the Cosmos world foundation model lets developers generate physics-accurate robot trajectories from a single image and language instruction. The GR00T-Dreams blueprint, announced at GTC March 2026, generates synthetic motion datasets without requiring any teleoperation data. Microsoft Azure and Nebius integrated NVIDIA’s Physical AI Data Factory blueprint, with FieldAI, Teradyne, and Hexagon Robotics already running on it.
The open-source community is the wildcard. Hugging Face’s LeRobot library has become the community standard for lightweight robot data recording and replay. NVIDIA’s Physical AI Open Datasets on Hugging Face have been downloaded over 4.8 million times. These datasets lower the floor for academic labs and startups, but they do not solve the quality problem. Roboflow’s active learning pipeline surfaces the issue directly: inconsistent labels early in the pipeline produce inconsistent behavior at deployment, and that is a hard problem to fix downstream.
Where the Money Goes Next
The real question for investors is not which approach wins in isolation. Foundation models need both real and synthetic data at different training stages: synthetic for variety and scale, real for dexterity and failure recovery. Goldman Sachs projects cumulative humanoid investment exceeding $50 billion by 2030. The percentage of that capital flowing to data infrastructure, currently a fraction, will have to catch up.
China is already moving. Max Fenkell of Scale AI told the House subcommittee on cybersecurity in 2026 that the U.S. is winning on AI model quality but losing on data and implementation, citing China’s strategy of funding mile-long warehouse facilities dedicated to gathering and labeling robot training data.
For founders building in this space, the structural advantage is provenance. The companies that maintain strict data lineage, covering who labeled what, under what task conditions, with what signal mix, own a moat that grows with every deployment. That is a harder asset to replicate than any model weight. The companies building that infrastructure, from industrial-scale annotation engines to biosignal-augmented ground truth platforms like Ground Truth Machine, are building what the physical AI stack cannot train without.
