In short
Nvidia’s Les Karpas will speak at TechCrunch Disrupt 2026 about why robotics still lacks a ChatGPT-like breakout moment. The core issue, he argues, is the absence of large-scale real-world data for physical AI.
- Nvidia’s Les Karpas will address the biggest bottleneck in robotics at TechCrunch Disrupt 2026.
- The central challenge for physical AI is the lack of internet-scale training data for real-world machines.
- Startups are trying to compensate with simulation, synthetic data, and foundation models across robot types.
- Nvidia sees robotics as a major AI frontier, with broad implications for manufacturing, mobility, and more.
Robotics has long promised a breakthrough like the one ChatGPT delivered for language AI, but that inflection point has not arrived yet. At TechCrunch Disrupt 2026, Nvidia Inception’s Global Head of Physical AI, Les Karpas, will argue that the biggest obstacle is not a lack of ambition or hardware, but the shortage of large-scale real-world data needed to train general-purpose robots.
Karpas is scheduled to appear on the Real World AI Stage at Disrupt, held October 13-15 at Moscone West in San Francisco, where he will discuss why physical AI has lagged behind software-first AI and what might finally push robotics into broader adoption. The session matters because it goes to the center of one of the industry’s most urgent questions: how can robots learn enough from the physical world to become truly useful at scale?
TechCrunch is positioning the conversation as part of a broader AI-focused program that will bring together founders, investors, and operators working across robotics, autonomous systems, manufacturing, mobility, and adjacent fields. For attendees, the appeal is not just a look at where the technology stands today, but a window into the bottlenecks that may determine which robotics companies survive the next wave of competition.
Why robotics is still waiting for its breakout moment
Robots have been in development for decades, and many are already performing valuable tasks in warehouses, factories, laboratories, hospitals, and defense settings. But the industry has not produced a consumer-facing leap comparable to ChatGPT’s rapid mainstream adoption after its launch in November 2022.
The reason, according to the framing behind Karpas’ session, is that robotics lacks the equivalent of the massive internet text corpus that helped large language models improve quickly. In language AI, developers could train systems on enormous volumes of digital content. In physical AI, the equivalent training material is far harder to capture because the real world is messy, variable, and expensive to model.
That gap matters because general-purpose robots must understand not only what to do, but how to move, balance, grasp, navigate, and recover from mistakes in unpredictable environments. Unlike text generation, where a model can learn from trillions of tokens, robots need data tied to movement, touch, timing, and spatial relationships.
What makes physical AI so difficult?
Physical AI is difficult because the data needed to train it is fragmented, costly to collect, and rarely standardized. A robot operating in a home, a warehouse, and a factory may face completely different surfaces, object shapes, lighting conditions, and safety constraints.
There is also a feedback problem. Software models can be refined using enormous online datasets, but robots typically require direct interaction with the world to improve, which makes scaling training much slower and more expensive.
That creates a chicken-and-egg dilemma for the sector:
- Robots need diverse data to learn general skills.
- Diverse data usually comes from large-scale deployments.
- Large-scale deployments are hard before robots are reliable enough to trust.
How are companies trying to solve the data problem?
They are turning to simulation, synthetic data, and foundation models designed to work across multiple robot types at once. The strategy is to create the scale that the physical world cannot provide cheaply enough on its own.
Simulation lets developers generate huge numbers of training scenarios without waiting for real-world experiences. Synthetic data can expose models to rare or dangerous situations. Foundation models, meanwhile, are intended to transfer knowledge between robots, tasks, and environments instead of training each machine from scratch.
Each approach is promising, but none has yet produced the kind of universal robotics platform that a ChatGPT-like moment would imply. The challenge is not simply making robots capable in controlled settings; it is making them adaptable enough to handle the unpredictable conditions that define real life.
| Area | Language AI | Physical AI / Robotics |
|---|---|---|
| Core training data | Internet text and images | Motion, sensor, and interaction data |
| Data scale | Massive and readily available | Limited, expensive, and fragmented |
| Training environment | Mostly digital | Real world plus simulation |
| Common bottleneck | Alignment and quality | Generalization and physical reliability |
| Path to scale | Large-model training and deployment | Simulation, synthetic data, and real-world fleets |
Who is Les Karpas, and why does Nvidia trust him to frame the debate?
Les Karpas is one of Nvidia’s key voices on physical AI because his role puts him at the intersection of robotics research, startup building, and industrial adoption. As Global Head of Physical AI at Nvidia Inception, he works with companies spanning robotics, automotive, manufacturing, mobility, and smart-city technologies.
That gives him a broad view of where the field is advancing and where it is stalling. The same startups trying to solve the data bottleneck are also the ones Nvidia interacts with most often through its accelerator and ecosystem programs.
Karpas also brings an unusually wide professional background to the stage. Over the course of his career, he has held roles connected to architecture, manufacturing engineering, startup leadership, venture investing, and corporate venture capital. His experience includes work with Stanley Black & Decker, Intellectual Ventures, Herman Miller, iRobot, and Cirque du Soleil.
That cross-disciplinary history matters in robotics, where technical breakthroughs are only part of the story. Commercial success often depends on manufacturing, safety, unit economics, deployment logistics, and customer trust.
Why Nvidia’s view carries weight
Nvidia has become one of the most influential companies in AI because its chips, software, and developer ecosystem underpin much of the industry’s training and inference infrastructure. In robotics, the company’s interest signals confidence that physical AI may eventually follow a similar scale-up path to language models.
Chief executive Jensen Huang has repeatedly highlighted robotics as a major growth area, and Nvidia’s continued investment has helped validate the field for startups and investors. When a company with Nvidia’s reach focuses on a segment, the market often pays attention.
Nvidia’s physical AI lead is expected to argue that robotics is not short on vision, but on the scale of data and deployment needed to unlock a broader leap forward.
Why Disrupt 2026 is using robotics to frame the next AI wave
TechCrunch Disrupt has long been designed as a place where founders can compare notes on what is working, what is overhyped, and where new businesses may emerge. In 2026, robotics fits neatly into that mission because it sits at the intersection of AI research and real-world industrial deployment.
The event’s Real World AI Stage is meant to showcase the practical obstacles facing systems that operate outside the screen. That includes robots, autonomous machines, and AI tools that need to function safely in physical environments rather than only in chat windows or software dashboards.
Karpas will not be speaking alone. He is expected to be joined by founders from Shield AI, Colossal Biosciences, FieldAI, and Foxglove, a lineup that reflects the range of physical-world AI challenges under discussion, from autonomous systems to infrastructure and specialized machine intelligence.
The broader message is clear: the next major AI growth stories may come not only from better text generation or coding tools, but from systems that can perceive, plan, and act in the physical world.
What the robotics sector still needs to prove
Robotics does not just need better models; it needs better business evidence. The sector must show that robots can be deployed profitably, maintained affordably, and improved continuously without requiring excessive human supervision.
That means solving several layers of the stack at once:
- Improving model training so robots can learn across tasks.
- Creating realistic simulation environments that transfer to the real world.
- Lowering deployment costs so fleets can expand faster.
- Building safety systems that satisfy enterprise and regulatory demands.
- Proving that customers will pay for automation that works reliably at scale.
The companies that crack those problems may shape the next decade of industrial AI. Those that fail may discover that impressive demos do not necessarily translate into durable products.
How the field differs from ChatGPT’s rise
The path to robotics adoption is likely to be slower and more sector-specific than the rise of ChatGPT because physical machines must operate under stricter constraints. A language model can make mistakes in a response and still be useful; a robot can create costly damage or safety issues if it misjudges a move.
That difference changes the economics of adoption. Many businesses may begin with narrow, repetitive use cases before expanding to more flexible machines. In that sense, the robotics industry may not need one giant consumer breakthrough as much as a series of practical wins across logistics, manufacturing, inspection, and specialized services.
Disrupt’s business pitch: insight, access, and dealmaking
TechCrunch is also using the Karpas session to underscore the conference’s value for networking and deal flow. The event will bring together more than 10,000 founders, operators, and investors, and organizers are emphasizing that the right conversations could lead to new partnerships, customers, and funding opportunities.
Ticket pricing is also part of the pitch. Attendees can save up to $200 before September 25, after which prices rise, and group buyers can save as much as 30%. For exhibitors, the last day to reserve a table is September 18.
Those details may sound promotional, but they are also a reminder of how the startup conference ecosystem works: timing matters, access matters, and the right stage conversation can become a catalyst for new business relationships.
Timeline: key milestones shaping the robotics debate
| Date | Milestone | Why it matters |
|---|---|---|
| November 2022 | ChatGPT debuts | Shows how a new interface can push AI into everyday use |
| 2020s | Robotics investment and experimentation accelerate | Startups seek practical uses for physical AI |
| September 18, 2026 | Last day to book an exhibit table | Deadline for companies seeking Expo Hall visibility |
| September 25, 2026 | Discount ticket window closes | Prices rise after this date |
| October 13-15, 2026 | TechCrunch Disrupt 2026 | Karpas appears on the Real World AI Stage in San Francisco |
What could come next for physical AI?
The most likely near-term outcome is not a single overnight transformation, but a gradual widening of capabilities. Robots are already useful in narrow, controlled settings. The question is whether new training methods and better data pipelines can make them more adaptive and less brittle in unfamiliar situations.
If the industry can do that, physical AI could follow a path similar to other major computing shifts: first as a niche capability, then as an essential infrastructure layer. That would affect logistics, manufacturing, healthcare support, retail operations, transportation, and even smart-city planning.
For now, though, the field remains in search of its defining leap. Karpas’ Disrupt session is likely to focus on why that leap has not arrived yet, what technical and commercial barriers still stand in the way, and why Nvidia believes the next major AI frontier may be the physical world rather than the screen.
In that sense, the session is more than a conference talk. It is a snapshot of an industry trying to answer one of AI’s biggest open questions: when will robots finally have their ChatGPT moment?
Frequently asked questions
What is physical AI?
Physical AI is AI designed to perceive and act in the real world through robots or other machines. Unlike software-only systems, it must handle movement, sensor input, safety constraints, and unpredictable environments, which makes training and deployment far more complex.
Why hasn’t robotics had a ChatGPT moment yet?
Robotics has not had a ChatGPT moment because it lacks the massive, standardized training datasets that helped language models scale quickly. Robots need large amounts of real-world interaction data, which is expensive and difficult to collect across many environments.
Who is Les Karpas at Nvidia?
Les Karpas is Nvidia Inception’s Global Head of Physical AI. He works with startups across robotics and related sectors, helping connect Nvidia’s ecosystem with companies building real-world AI systems.
When is TechCrunch Disrupt 2026?
TechCrunch Disrupt 2026 runs from October 13 to October 15 at Moscone West in San Francisco. Karpas is scheduled to appear on the Real World AI Stage during the event.
How are robotics companies trying to overcome the data shortage?
Robotics companies are using simulation, synthetic data, and foundation models trained across multiple robot forms. These methods are meant to create scale without waiting for enough real-world deployments to generate sufficient training data.
