What I Look for When Investing in Physical AI

Physical AI is one of the most exciting areas in venture right now.

That is exactly why investors have to be more disciplined, not less.

A robot walking across a stage is memorable.

A machine doing economically useful work every day is investable.

Those two things overlap less often than the headlines suggest.

At Nirman Ventures, I spend most of my time around AI and robotics for construction, manufacturing, logistics, infrastructure and other parts of the physical economy.

When I look at a company in this category, I keep coming back to a handful of questions.

Is the workflow painful enough?

I start with the problem, not the robot.

What is happening today that is expensive, dangerous, repetitive, slow or constrained by labor?

And is it painful enough that a customer will actually change behavior to solve it?

There are many workflows people complain about and very few they will allocate budget, implementation time and organizational attention to fixing.

I like problems where the value is already visible in the customer's operations: labor, rework, throughput, downtime, safety, cycle time, material movement, inspection and utilization.

The strongest markets do not need to be convinced that the problem exists.

Is the first wedge narrow enough?

Physical AI founders are often ambitious by nature.

That is a strength until it makes the first product too broad.

Automate the warehouse. Replace construction labor. Build the general-purpose industrial robot.

Those may be long-term visions. They are rarely good first products.

I prefer a company that can identify one high-value task, solve it exceptionally well and then expand.

The wedge should be narrow enough to deploy, but valuable enough to matter.

Earn trust on one task. Expand after you have earned the right.

Does the technical advantage matter in the field?

Physical AI companies can show impressive benchmark results that never become meaningful customer advantages.

I care about the technical moat, but I care even more about where it shows up.

Does better perception reduce interventions? Does a stronger autonomy stack allow deployment in environments competitors cannot handle? Does hardware design improve reliability or serviceability? Does the system recover from edge cases instead of stopping?

A technical advantage becomes a business advantage only when the customer can feel it.

Is the customer buying an outcome or admiring a demo?

There is a lot of innovation theater in emerging technology.

Customers want to show they are experimenting. Startups want logos. Investors want validation.

Everybody can leave the pilot happy without a production contract ever happening.

So I want to know what success means before the deployment starts.

What metric moves? Who owns the budget? What happens if the pilot works? How many additional units or sites could follow?

Interest is not adoption. A pilot is not scale.

Does deployment get easier?

This may be the single most important operating question in Physical AI.

The machine has to arrive. It has to be installed. It has to be integrated. It has to survive the environment. Someone has to support it.

So I look at deployment velocity.

How many days from arrival to productive operation? How many engineers are involved? How much site-specific work is required? Can installation eventually be handled by customers or channel partners? Can the system diagnose itself remotely?

Every deployment should teach the company how to make the next deployment cheaper and faster.

If that learning is not happening, growth can actually make the business harder to operate.

Do the unit economics include reality?

One of the easiest mistakes in hardware investing is to look at machine margin and miss the cost of delivering the service around it.

Installation. Field support. Travel. Replacements. Maintenance. Remote monitoring. Spare parts. Customer-specific engineering.

I want to see the economics of the complete outcome.

That does not mean early margins need to be beautiful. They often will not be.

What matters is whether there is a credible path for them to improve.

Volume can lower hardware costs. Autonomy can reduce human support. Better reliability can reduce service costs. Standardized deployment can reduce implementation expense.

The question is not whether margins are good today. It is what gets structurally better at scale.

Does the team respect the operator?

This one is personal for me.

I spent more than two decades as an operator in construction, engineering and real assets before becoming an investor.

I have seen brilliant technology fail because the people building it assumed the customer simply needed to be educated.

Sometimes the customer does need education.

Often the product needs humility.

Operators understand constraints that are invisible from a lab: safety rules, workflow dependencies, procurement cycles, maintenance realities, liability, uptime requirements and the cost of stopping production.

The founders I like do not treat that domain knowledge as resistance.

They treat it as data.

Respect for the operator is not a soft skill. In Physical AI, it is part of product development.

Is there a moat beyond the hardware?

Hardware can absolutely be defensible.

But I prefer when the moat compounds across several layers.

Data from deployment. Proprietary autonomy. Workflow integrations. Fleet learning. Service infrastructure. Distribution. Customer trust. Deployment playbooks. Safety certification. Manufacturing know-how.

The strongest Physical AI businesses can become harder to replace every month they operate.

I want to know what improves with every unit in the field.

Is the form factor solving the problem or selling the story?

Humanoids attract attention for good reason.

The world was built around human dimensions, human tools and human movement.

There are workflows where that flexibility will matter enormously.

But I do not assume humanoid is the answer simply because it is the most exciting form factor.

A wheeled platform may be cheaper. A fixed arm may be more precise. An autonomous piece of equipment may already have the mechanical advantage required.

The machine should fit the workflow, not the fundraising narrative.

Can this become a large company?

Finally, venture math still matters.

A wonderful product in a tiny workflow may create a good business and still not create a venture-scale outcome.

So after understanding the wedge, I want to understand the expansion path.

More tasks. More sites. More industries. More value captured per machine. More software. More fleet intelligence. More workflow ownership.

The best companies can start narrow without staying small.

What I am ultimately underwriting

I do not invest in robots because robots are interesting.

I invest when a machine can solve a real operating problem in a way that becomes more reliable, more economical and more valuable as it scales.

That requires deep technology.

It also requires something the market talks about less: discipline.

Discipline about the workflow. Discipline about deployment. Discipline about economics. Discipline about what not to automate yet.

Physical AI may become one of the defining technology cycles of the next decade.

But the companies that win will not be the ones with the best demo.

They will be the ones that turn intelligence into repeatable economic value in the real world.