Software AI may become a commodity. Physical intelligence may not.

It’s one thing to build a model that can answer questions or write code. It’s another to
build one that can see a messy workplace, understand a task, move safely, handle fragile
objects, and recover when something goes wrong.

Adam Rogers’s recent Scientific American article, Grasping the World, looks at Google
DeepMind’s work on Gemini Robotics. DeepMind isn’t trying to teach one robot one job.
It’s trying to build a more general intelligence that can use what it learns across
different tasks and different robot bodies.

Berkshire’s investment in Alphabet gives its shareholders a small stake in that
possibility. Robotics is tiny compared with Alphabet’s current businesses, so I
wouldn’t make it part of the investment thesis today. But it could matter later if
physical AI becomes a large market.

Rogers gets at the hard part. Gemini already knows a lot about the world from text and
images. But knowing what a wine glass is doesn’t tell a robot how tightly to hold it.
DeepMind is trying to teach Gemini how to use what it knows to control a physical
machine. One of the robots it works with is Apptronik’s Apollo humanoid. In one
difficult dustpan test, Apollo succeeded only 32% of the time. The task sounds trivial,
but that’s the point: everyday physical work demands a kind of touch and control that
software AI doesn’t yet have.

The data problem may be where things get more interesting. The Internet has endless
text, images, video, and code, but far less data about touch, pressure, force, grip, or
body position. Some of that gap can be filled with simulation and videos of people
doing tasks. But real robots working in the physical world may still produce especially
valuable data.

That suggests a possible flywheel: more robots create more useful physical data, which
improves the AI, which makes the robots better.

Of course, that flywheel depends on lessons learned by one kind of robot carrying over
to others. That’s a DeepMind goal, not a proven fact.

Tesla is taking the vertical route: build the body, software, AI, training system, and
manufacturing process together. Google is making a different bet: build the intelligence
layer rather than the whole machine. Its work with Boston Dynamics is one example.
Nvidia, startups, and Chinese companies are chasing similar ideas.

But Google’s approach raises an important problem. If other companies own the robots,
who owns the data they produce? If Google doesn’t get enough of it, the data advantage
could end up with the robot makers instead.

Tesla avoids that problem. It owns its robots, so it owns their data. The tradeoff is
that it also has to build and sell all of those robots itself.

So the real question may be:

If a moat develops in physical AI, could it come from owning the machine, owning the
data, or owning the learning system that gets better across many machines? My take
is that the data and learning loop could prove more valuable than the robots
themselves.

Morgan Stanley estimates a humanoid market of more than $5 trillion by 2050, based on
roughly a billion robots in service and the replacement value of units sold into that
installed base. That’s a forecast, not a fact, but it shows why this is more than an
experiment.

I’d watch for paid deployments, wider Gemini use across independent robot makers, and
signs that Alphabet can turn that use into revenue.

It’s far too early to call any of this a moat. But we may be starting to see where
one could form.