Remember when “AI” meant a chat window and a blinking cursor? For a long stretch, nearly every tool that crossed my desk here at agntbox was some flavor of text in, text out. Summarizers, code helpers, email drafters. Some were useful and some were filler. But all of them lived behind glass, and none of them ever touched anything you could hold.
That era isn’t over, but it has company now. The phrase landing in my inbox more than any other lately is “physical AI,” and most of those pitches arrive with a shiny humanoid robot in the hero image. So it was refreshing to see someone with real inside experience push back on that framing.
Who’s talking, and why I’m listening
Amulya Vishwanath spent seven years leading product marketing for NVIDIA’s robotics platforms. She’s now on the investing side, putting money behind physical AI companies. That combination gets my attention. Someone who marketed the platform knows exactly where the polish ends and the rough edges begin. Someone writing checks has to care about those rough edges, because that’s where money gets lost.
Her core argument is that physical AI is a much wider category than robots. In her words, it can be “a molecule, a material, a medicine,” or a machine on a factory line. That’s a big tent. It puts drug discovery and materials science in the same conversation as warehouse automation and assembly lines.
I think she’s right, and I think it matters for anyone evaluating tools in this space. If you only picture a two-legged robot carrying boxes, you’ll miss most of where this technology is likely to show up first. The less photogenic applications tend to be the ones that quietly ship.
The “Lego era” is a useful mental model
Vishwanath also describes robotics as being in its “Lego era,” with basic building blocks like actuators now available as pieces you can snap together. As someone who reviews toolkits for a living, this is the part of her pitch I find most practical.
It describes a market where the components exist but most of the finished sets are still in the box. That cuts both ways for builders and buyers. Modularity means you aren’t locked into one vendor’s full stack from day one, and you can swap parts as better ones arrive. It also means integration becomes your problem. Anyone who has tried to make three “compatible” SDKs actually talk to each other knows that’s where weekends disappear.
So when a vendor tells me their physical AI product “just works,” the Lego framing gives me a better question to ask. Which bricks did you build, and which ones are you expecting me to supply?
Nvidia’s numbers aren’t small
This isn’t a side project for the company Vishwanath came from. Nvidia CEO Jensen Huang has described the physical AI business as already driving nearly $10 billion in annual revenue. That’s a real business, not a slide in a keynote.
The company has kept pressing. At GTC 2026, Nvidia doubled down on physical AI and humanoids. It has also been broadening its push with robotics updates, foundation models, and edge hardware, filling in more of the ecosystem around the chips. Huang also framed a joint effort with T-Mobile and Nokia as turning the 5G network into a distributed AI computer, which hints at where some of this processing could eventually live.
The piece I care most about as a reviewer is safety. Nvidia has expanded its safety offering for robotics, and robotics companies are using it for safer deployments of autonomous mobile robots and humanoid robots. Machines that move around people need a much higher bar than a chatbot that occasionally gets a fact wrong. Seeing safety treated as a product line instead of a footnote is a good sign.
What works, what doesn’t (so far)
Based on what’s been shared, here’s my honest scorecard:
- Works: The broader definition. Treating molecules, materials, and medicine as physical AI pulls attention toward applications that may deliver value sooner than humanoids.
- Works: The safety focus. Tooling aimed at safer deployment of mobile and humanoid robots addresses the biggest real-world blocker.
- Works: Ecosystem breadth. Foundation models plus edge hardware plus network partnerships means fewer gaps for builders to fill alone.
- Doesn’t (yet): The integration burden. A “Lego era” means someone still has to assemble the set, and that cost rarely shows up in a demo.
- Watch out: The source of the optimism. The loudest voices here are Nvidia and people who built Nvidia’s robotics marketing. Their insight is valuable, but they’re naturally bullish.
My take
The most useful thing Vishwanath said isn’t about robots at all. It’s the reminder that “physical” is a much bigger word than “humanoid.” For anyone shopping for tools, that shift changes what you should be testing. Ask how a platform handles simulation, safety, and messy real-world data, not just whether the robot in the video can wave.
I’ll keep reviewing these toolkits the way I review everything else. Does it do what it says, how much glue do I have to write, and would I trust it near people? The bricks are on the table now. I’m more interested in who builds something that holds together.
🕒 Published: