Physical AI is more than robots, says the Nvidia veteran now investing in it

The Lego Problem

Everyone wants to believe we are one software update away from a C-3PO future. If you follow the venture capital trail or the latest AI demos, the narrative is simple: we have the brains now, so the bodies will follow effortlessly. But if you talk to people who actually spent years inside the belly of the beast—places like Nvidia—the story gets a lot more grounded.

Amulya Vishwanath, a former product marketing lead at Nvidia who now spends her time investing in the space, describes our current moment as the Lego era of robotics. It is a sobering comparison. When you think of Legos, you think of modularity, but you also think of limitations. You are clicking together pre-made blocks to see what sticks. We are not in the era of seamless, biological-grade automation yet. We are still trying to figure out if the hand can actually hold the screwdriver without breaking itself or the tool.

For founders building in the physical AI space, this perspective is vital. The industry is currently obsessed with general foundation models—the idea that if you feed enough video data into a transformer, a robot will magically learn how to fold laundry. While the software side is moving at light speed, the physical hardware and the standards governing it are stuck in a crawl.

Beyond the Humanoid Hype

We see a lot of videos of humanoid robots walking through forests or making coffee. They look cool, and they sell well on social media. But from a builder’s perspective, humanoids are the hardest possible starting point. Vishwanath’s point is that physical AI is much broader than just making machines that look like us. It is about intelligence interacting with the physical world in any form factor.

The bottleneck isn't just the AI brain; it’s the actuators, the sensors, and the tactile feedback systems. In the digital world, if your code fails, you get an error message and you try again. In the physical world, if your code fails, your $100,000 prototype walks into a wall and shatters. The cost of failure in physical AI is orders of magnitude higher than in pure software, and the building blocks—those Lego pieces—are still incredibly fragile.

The Safety Gap

One thing that often gets lost in the rush to market is safety. In a factory setting, robots have historically been kept in cages. They are powerful, dumb, and dangerous. Physical AI aims to take them out of the cages so they can work alongside humans. But we don't have the universal safety standards for that yet.

If you are building a startup in this space, you can't just move fast and break things. If you break a human, your company is over. This is why the Lego era is so frustrating for investors and founders alike. We are rebuilding the basic infrastructure of how machines move and sense their surroundings. We are still arguing over how a robot hand should even be designed. Should it have five fingers? Three? Should it use suction?

The Founder Perspective: Don't Wait for the General Model

There is a trap right now where founders think they can just wait for a GPT-level model for robotics to solve their problems. They assume that once the model is ready, they can just plug it into any off-the-shelf hardware. This is a mistake. The integration of software and hardware in physical AI is not a plug-and-play situation.

Real success in the current market is coming from those who are solving specific, narrow physical problems rather than trying to build a general-purpose butler. Builders need to focus on the boring stuff: reliability, durability, and specific utility. If you can make a robot that does one task with 99.9% reliability, you have a business. If you make a humanoid that does twenty tasks with 70% reliability, you have a science project.

Vishwanath’s shift from Nvidia to the investment side suggests a hunt for the companies that understand these mechanical nuances. The winners won't just be the best coders; they will be the people who understand how to harden hardware for the real world. They are the ones who realize that the Lego blocks we have right now aren't good enough, and they are busy inventing better ones.

The Reality of the Roadmap

We are seeing a massive influx of talent moving from pure AI into robotics, and that’s great for the ecosystem. But these engineers are quickly realizing that the physical world is messy. It has friction, gravity, and unpredictable lighting. You can't simulate every variable, no matter how good your digital twin technology is.

This means the timeline for a truly autonomous physical world is longer than the hype cycles suggest. We are currently in the infrastructure phase. We are building the roads and the traffic lights. The cars—the truly intelligent, autonomous robots—will come later. For now, the opportunity lies in the tools, the specialized sensors, and the niche applications that prove the tech works in controlled environments.

The physical world doesn't have an undo button. Builders who respect the hardware will outlast the ones who think it's just a container for their code.

We need to stop looking at robotics as a solved hardware problem waiting for a software solution. It is a holistic engineering challenge. If you are entering this space, prepare for a marathon, not a sprint. The Lego era is a time for experimentation, but it is also a time for rigorous testing and humble expectations about what these machines can actually do today.

Takeaway for Builders

Stop chasing the humanoid dream unless you have a decade of runway and a team of mechanical geniuses. The immediate value in physical AI is in the components and the specialized applications. Focus on building the specific blocks that the rest of the industry is missing. The world doesn't need another shaky robot video; it needs a robot that doesn't break when it touches a door handle. Solve the physical reliability problem, and the software will finally have a body worth inhabiting.

Read the original at TechCrunch Startups →