Three Things That Matter in Physical AI Right Now

At the Nebius Physical AI Awards I sat in on a VC panel — brilliantly moderated by Anubhav Maheshwari — about what actually matters for teams building in physical AI right now. Three themes stuck with me.

1. Synthetic data and sim are the quiet unlock

The real bottleneck isn't building the first five robots. It's going from a cool demo plus one or two pilots to hundreds of robots across many customers.

That's where better simulation and synthetic data start to matter more than another “wow” video:

• Generalising across environments without collecting data from scratch at every site
• Shortening the 0 → 1 → 10 → 100 rollout curve
• Tightening the loop between research labs and deployment teams

The companies that treat sim and synthetic data as a core go-to-market enabler — not a side-quest for the research team — will scale faster than those trying to brute-force every new customer with field data.

2. Foundation models for robotics: underfunded and overfunded

One line from the panel I really liked:

The most underfunded and overfunded category in robotics is the same one: foundation models for robotics.

And it's true:

• Underfunded if you look at how central these models will be to the future robotics stack: perception, planning, manipulation, multi-task behaviour.
• Overfunded if you look at how many teams are trying to raise on a polished video and a “we’ll figure out the business model later.”

If you're building here, the bar is moving from “cool model on a robot” to:

• Who finances the fleets?
• What do unit economics look like at scale, not just in a pilot?
• How do you sell, deploy, maintain and upgrade thousands of systems, not ten?

The model is necessary, but not sufficient. The value is in the system and the business around it.

3. The biggest wins are likely in “unsexy” domains

Humanoids and sci-fi general-purpose robots will always attract attention (and capital). But a lot of the real value will be created in places that:

• Don't look glamorous on stage
• Replace dangerous, repetitive, or deeply annoying work
• Sit in industries most VCs don't think about until someone shows them a working robot and real revenue

Think inspection, maintenance, logistics edge cases, very specific manufacturing operations — domains where:

• The pain is obvious
• The budget already exists
• Safety and compliance give you a strong pull, not just a “nice to have”

These “niche” robots can end up being category kings with very defensible businesses.

The takeaway

My own mental note after the panel: over the next few years, the real differentiation in robotics and physical AI won't just be who has the flashiest humanoid. It will be who can combine:

• Synthetic data and sim → fast generalisation and rollout
• Serious foundation models → real multi-task capability, not just a demo
• “Boring” but painful verticals → obvious ROI, strong pull from customers