Robots are waiting for a ChatGPT moment: Nvidia’s Les Karpas explains why at TechCrunch Disrupt 2026

We have all spent the last few years watching LLMs get faster, smarter, and eerily more human. In the digital world, we have hit a velocity that feels uncontrollable. But when you step outside and look at the physical world, the revolution looks a lot more like a slow crawl. We were promised C-3PO; we got a vacuum cleaner that occasionally gets stuck on a rug. At TechCrunch Disrupt 2026, Les Karpas from Nvidia laid out a reality check for why the robotics industry hasn't had its GPT-3 moment yet.

The Barrier Between Bits and Atoms

The problem is the fundamental difference between processing text and processing physics. When an LLM fails, it hallucinates a fact. When a robot fails, it breaks a window, ruins a piece of hardware, or endangers a human. Karpas pointed out that while we have massive datasets for human language, we lack the same depth of data for high-fidelity physical interaction. We are trying to teach machines to dance before they truly understand gravity.

For builders, this is the core bottleneck. Digital AI scales because the environment is standardized. A Python script runs the same on my machine as it does on yours. But in robotics, every environment is a specialized edge case. The lighting changes, the floor texture varies, and the objects the robot needs to interact with are never in the exact same spot twice. This lack of standardization is what keeps robots trapped in factory cages rather than roaming our homes.

Why Generative AI Hasn't Solved Hardware Yet

There is a common misconception in the founder community right now: that if we just throw more compute and more tokens at robotics, the 'physical' problem will solve itself. Karpas effectively debunked this hope. Deep learning is great for recognition and pathfinding, but it struggles with the tactile feedback loops required for fine motor skills. A robot needs to 'feel' how much pressure to apply to an egg versus a hammer. We haven't figured out how to digitize that sense of touch at scale.

Nvidia is positioned as the picks-and-shovels provider for this space, mostly through their simulation platforms. They are betting heavily on the idea that we can train robots in digital twins—simulated worlds where we can run a thousand years of 'experience' in a few hours. If a robot crashes in a simulation, it costs nothing. But even Karpas admits there is a 'reality gap.' The transition from a perfectly rendered digital world to the messy, dusty, hardware-vibrating real world is where most startups die.

The Founder Strategy: Don't Chase the Generalist Dream

Middle-market founders and builders often get lured into the idea of building a general-purpose humanoid. It’s a great story for a pitch deck, but it’s a nightmare for an engineering roadmap. The takeaway from the current state of the industry is that the winners won't be the ones building 'The Robot.' They will be the ones building the vertical middleware that solves one specific physical problem perfectly.

• Focus on simulation-to-reality (Sim2Real) pipelines: If you can’t get your data out of the lab and into a noisy environment, you don't have a product.
• Prioritize durable hardware: software is easy to patch; a bent actuator is a week of downtime.
• Solve for 'Dumb' problems first: The industry is waiting for a brain, but it also needs better nervous systems.

We are currently in the 'dial-up' phase of physical AI. We have the connectivity and the basic ideas, but the bandwidth between the computer's intent and the robot's action is still far too thin. Karpas’s perspective suggests that Nvidia is ready to provide the compute, but the industry still needs a breakthrough in how we collect and tokenize physical interaction data.

The Skeptic’s View on the Timeline

If you listen to the hype, we are eighteen months away from robots folding our laundry. If you listen to builders who are actually in the trenches, we are still struggling with basic battery life and joint friction. The 'ChatGPT moment' for robotics won't happen because of a better chatbot interface. It will happen when we find a way to map the chaos of the physical world into a language that neural networks can actually comprehend without exploding.

The gap between a robot that works in a controlled demo and a robot that works in a living room is wider than the gap between a calculator and a supercomputer.

As builders, our job isn't to wait for the breakthrough. It's to build the infrastructure that makes the breakthrough inevitable. Right now, that means obsessing over data quality, sensor fusion, and making sure our hardware can survive more than ten minutes of actual labor. The hardware revolution is coming, but it won't be as clean or as fast as the software one was.

Read the original at TechCrunch Startups →