According to Crunchbase data, venture capital in robotics and physical AI passed $14 billion in the first two months of 2026, matching all of 2025 in 60 days. March added another $2.25 billion across five deals. The capital isn’t just growing. It’s accelerating. This acceleration sits against the backdrop of $700 billion in cloud infrastructure spending from the four largest hyperscalers in 2026 alone.
But the money isn’t flowing where most people assume. The companies raising the largest rounds aren’t hardware leaders. They’re data companies that happen to build robots.
The data moat, not the hardware moat
Mind Robotics raised $500 million in a Series A co-led by Accel and Andreessen Horowitz. The company was spun out of Rivian in November 2025 by founder RJ Scaringe. Its advantage isn’t mechanical engineering. It’s Rivian’s factory floor, where every pick, place, error, and recovery is captured as training data. The robots learn from real production, not simulated environments.
Skild AI raised $1.4 billion at a $14 billion valuation and is already deploying its “universal robot brain” inside Foxconn’s Houston factory, where NVIDIA’s Blackwell GPU racks are built. Apptronik’s Apollo humanoid is being tested at Mercedes-Benz factories and GXO Logistics warehouses. These aren’t lab demos. They’re production deployments.
The pattern across all of them: proprietary operational data is the competitive advantage. The hardware is converging. The training data isn’t.
The counter-thesis worth watching
Yann LeCun, Turing Award winner and former Meta Chief AI Scientist, raised $1.03 billion for AMI Labs in what became Europe’s largest seed round. His argument: LLMs are fundamentally wrong for robotics. Language models predict the next word. Robots need to predict the next physical state of the world. LeCun’s JEPA architecture works in compressed representation space, closer to how human brains actually process physical environments.
If he’s right, the companies that dominate physical AI won’t be the ones that scaled language models. They’ll be the ones that built world models. It’s too early to call, but $1.03 billion from Bezos, NVIDIA, Samsung, and Toyota says the bet is serious.
The skeptics have a point, and it matters
Rodney Brooks, co-founder of iRobot and one of robotics’ most respected voices, expects a significant hype cycle followed by disappointment. He points out that human hands have roughly 17,000 specialized touch receptors. No robot comes close. Integration costs can exceed $500,000 per production line. Safety certification takes 12-18 months per model.
The failures are real too. Zume Pizza raised $500 million on robotic pizza delivery and collapsed. iRobot itself recently filed for bankruptcy. Technical capability and commercial viability are not the same thing.
The skeptics’ timeline is probably right for general-purpose humanoids in unstructured environments. But the narrow industrial use cases already work. Structured environments, repetitive tasks, positive ROI within 18 months. Goldman Sachs projects humanoid robots at $15,000-$20,000 per unit at scale, with shipments reaching 50,000-100,000 units in 2026. The scope will expand from structured to unstructured, but the economics are already proven in factories and warehouses. All of it, though, depends on the same physical infrastructure chokepoints that constrain every AI deployment: chips, energy, and data center capacity.
What this means if you run physical operations
The 2.1 million manufacturing jobs projected to go unfilled by 2030 aren’t coming back as human hires. The question for most companies isn’t whether to buy robots. It’s whether to become a trainer or just a buyer.
Companies that start documenting their operational knowledge now, the workflow sequences, the maintenance edge cases, the quality inspection patterns, won’t just adapt when physical AI becomes affordable. They’ll be positioned to feed their specific institutional knowledge into systems that learn from it. The data is the moat.
Jensen Huang said it at GTC: “The ChatGPT moment for robotics is here.” The moment may be arriving. Whether your operational data is ready to meet it is a decision you can make now.
Related: You’re Not Using AI Wrong. You’re Building Wrong. explores the same organizational readiness gap in software AI: the tools work, but the redesign is the actual work.
Update: The Physical World Underneath AI maps the four physical chokepoints that all this capital assumes will keep working.
Questions this article gets
How much venture capital has gone into physical AI and robotics in 2026?
According to Crunchbase data, venture capital in robotics and physical AI passed $14 billion in the first two months of 2026, matching all of 2025 in 60 days. March 2026 added another $2.25 billion across five deals.
Why is operational data more important than hardware in physical AI?
The companies raising the largest funding rounds are those with proprietary operational data, not the best hardware. Mind Robotics' advantage comes from Rivian's factory floor where every movement becomes training data. Hardware is converging across companies. Training data is not.
What is Yann LeCun's counter-thesis on robotics and LLMs?
Yann LeCun argues that LLMs are fundamentally wrong for robotics because language models predict the next word while robots need to predict the next physical state of the world. His JEPA architecture works in compressed representation space, closer to how human brains process physical environments. His new company AMI Labs raised $1.03 billion in Europe's largest seed round.