The Most Important Number in Physical AI Is 42

By Lior Handelsman

Why the funding data says robotics is underbuilt rather than overheated, and what now separates a defensible Physical AI company from one standing on the same open platform as everyone else.

• Headline robotics rounds look like a bubble, but only 42 robotics companies have raised more than $30 million in the past five years, against 745 software companies over the same period.
• NVIDIA has released the model layer, the humanoid control stack, and the on-robot compute as an open platform, which moves the durable advantage away from model access.
• Robotics has roughly 300,000 hours of manipulation data in total, so proprietary data and a closed loop between real deployment and simulation are now the scarce assets.
• Israel holds one of the largest computer vision concentrations outside the US and China, but almost all of its 125 Physical AI companies stop at perception rather than action.

Almost every conversation about Physical AI eventually pops up the same question: are we in a bubble? It’s hard to blame anyone for asking. Figure raised a billion dollars at a $39 billion valuation. Skild raised $1.4 billion for a robotics foundation model; and Travis Kalanick, former CEO of Uber, raised $1.7 billion for a new robotics company. The numbers look like it’s 2021.

But anyone who stops at the headlines misses the number that actually matters. Over the past five years, 745 software companies raised more than $30 million. In robotics: 42.

Are We in a Physical AI Bubble?

Eighteen times fewer funded companies – against a base market larger than any software market I know. So the giant rounds in the headlines are capital catching up on a decade of lag, in a category that suffers from a deep shortage of good companies.

What Changed in 2026 – Did NVIDIA Just Remove Robotics’ Biggest Barrier?

What makes that shortage especially interesting is that this year, the single biggest barrier disappeared. In 2026 NVIDIA released the brain as an open platform: Cosmos 3, an open model unifying text, image, video, sound, and action. GR00T, for full humanoid control. And the Jetson Thor chip, which runs all of it on the robot itself, for $1,999. Today robotics is the fastest-growing category on Hugging Face. I’ve been investing in this field for nine years, and I learned the lesson of this shift back at SolarEdge, when my competitor was Chinese hardware priced at zero. When the brain becomes a commodity, value moves. To data, to simulation, to integration.

How Reliable Are Humanoid Robots Today? The GPT-2.5 Moment

Before we go on, a few words of honesty. Robotics is in its GPT-2.5 moment: the capability is already here, and the reliability is still on its way. Figure’s humanoids handled more than 90,000 parts at a BMW plant – production work by any definition. And still, most humanoid robots in the world are in a pilot stage, and even DeepMind’s most advanced model succeeds somewhere between 32% and 92% on complex contact-rich tasks. The road from 80% success to 99.9% is far from linear, because in the physical world there is no Undo. There is no “you’re right, let’s try again.” The robot walks into the table. The drone flies into the wall. The autonomous vehicle hits a pedestrian.

Why Is There So Little Robotics Training Data?

Closing that gap takes data, in volumes this field has never seen. And that’s where the category’s real problem lies – and its real opportunity. The digital world had trillions of words and a billion hours of video to train on. The physical world has, in total, roughly 300,000 hours of robotic manipulation data. There is no open dataset for friction, for heat, for wear, for gravity. There is no Wikipedia for factory-floor physics.

Physical data is the new steel. Hard to produce, even harder to copy.

The only way I know to produce it at pace is a closed loop. You start with real data from the field, and use it to calibrate a simulation that knows real physics: the friction of the wheel, the wear of the milling tool, the torque of the screwdriver. The simulation generates training volume that can’t be recorded in reality. The volume trains the model. The model goes out to the field, meets reality, and returns new data that improves the simulation.In an era when everyone stands on the same open model platform, that loop is a significant moat.

Why Does Deployment Speed Decide Which Robotics Companies Win?

Alongside the loop there’s a second question that decides companies: deployment speed. About 75% of the cost of an automation project sits in setup, integration, and re-engineering – and reliability and unit economics are only proven in the field, at a paying customer. So my advice to founders is simple. Don’t go to markets with heavy regulation and endless adoption cycles. Go where the pain is real and the willingness to try already exists.

Israel is holding a card most of the market hasn’t priced in yet. Israel already owns the perception layer: more than 60 computer vision companies with over 2,300 patents, one of the largest concentrations in the world outside the US and China. But when you look at the Israeli Physical AI map – 125 companies – you see that almost all of them stop at perception. The big prize is turning perception into action, and that layer is wide open. And the talent from the military units and the ecosystem, raised on sensing, optics, sensor fusion, and autonomous systems, is exactly the skill mix this field demands.

What Four Questions Should Every Physical AI Founder Be Able to Answer?

When a company in this space comes to me, I ask four questions:

1. What is the physics insight that gives you an unfair advantage – the kind that comes from depth in the domain, because access to the model is something everyone has today?
2. What proprietary loop do you have between real data and simulation?
3. Is the brain your product or your infrastructure, because competing with NVIDIA is a capital game of a different order of magnitude?
4. How fast will you get to real deployment at a paying customer?

The founder who can address these questions well, is playing a different game than Figure and Skild. A game where 42 companies split a market that 745 were supposed to be fighting over.

We are in the years when this game is still open.