TechCrunch Disrupt 2026 splits AI in two, and the real-world half is where the hard problems live

TechCrunch Disrupt 2026 is adding a second AI stage, and that says more about the industry than any keynote could.

The original AI Stage stays focused on software, security gaps, and the business-model churn AI is forcing on SaaS. The new Real World AI Stage covers the messier work: robots, autonomous hardware, edge deployments, defense systems, industrial machines, and the biology-adjacent stuff that still makes tech people argue over drinks.

The split makes sense. AI has already broken into two very different problem sets. One is about models, workflows, and software economics. The other is about physics, sensors, latency, safety, and the fact that the real world doesn’t reset when a system gets something wrong.

The new stage runs October 13 to 15 at Moscone West in San Francisco, alongside the rest of Disrupt 2026. The lineup leans hard toward companies trying to make AI do something outside a browser tab. That’s the useful signal.

Why physical AI gets its own stage

A lot of AI conferences still treat robotics and autonomous systems like a side topic. That’s lazy. The stack is different enough to deserve its own room.

Large language models had the internet. Self-driving systems had millions of hours of road data. Robots, in most categories, have neither that scale of data nor the clean feedback loops that made modern AI take off in the first place. The result is a stubborn gap between demo-ready robotics and something you can ship without crossing your fingers.

That gap shows up fast:

• Data scarcity. You can scrape text at scale. You can’t scrape real-world manipulation data as easily.
• Sim-to-real failure. A model trained in simulation can look solid until lighting changes, surfaces vary, or a motor starts to wear out.
• Latency and connectivity. Many deployments can’t rely on cloud round trips.
• Safety and certification. A bad autocomplete is annoying. A bad control loop can be catastrophic.
• Operations. Hardware has supply chains, maintenance cycles, and field failure modes. Software teams often learn that the hard way.

That’s the thread running through the new stage. TechCrunch isn’t pretending physical AI is just another model wrapper. It’s treating it like a systems problem.

The robot data problem is still the bottleneck

The session likely to draw the most attention is the one asking when robots get their ChatGPT moment. That’s the right question, but it comes with a catch: physical AI won’t follow the LLM playbook cleanly.

For text models, the scaling story was blunt. More data, more compute, better token prediction, then better capability. Robotics and embodied systems need more than that. They need interaction data. They need perception tied to action. They need policies that can handle uncertainty, partial observability, and objects that move when you touch them.

That’s why Nvidia’s Les Karpas, head of physical AI, matters here. Nvidia has been pushing simulation, foundation models for robotics, and synthetic data tooling. That helps, but it also points to the core limitation: synthetic data goes only so far before reality starts breaking your assumptions.

For engineers, the question isn’t whether physical AI can work in theory. It’s whether the data pipeline, model design, and deployment environment line up well enough to make it reliable enough to matter.

That’s still the bar.

Safety is the product

One of the stronger sessions on the stage is the one on building AI systems when failure is not an option. That’s the session most startup decks quietly avoid.

When AI touches aircraft, vehicles, defense systems, or industrial machinery, “move fast” becomes a problem. A bad output can ground a fleet, damage equipment, or put people at risk. The product discipline changes completely.

Shield AI’s Nate Michael is slated to speak on this, which fits. In hard-tech AI, safety isn’t something you bolt on later. It shapes architecture from the start. You need validation frameworks, test coverage that goes beyond unit tests, and operational controls that look a lot more like systems engineering than app development.

The questions get practical fast:

• What does failure look like in the field?
• How do you detect model drift when the environment changes?
• What level of autonomy is acceptable at each deployment phase?
• What happens when the model is uncertain?
• Which parts of the system stay deterministic?

For teams building in defense, aviation, or industrial automation, those questions matter more than benchmark bragging rights. A model can be state of the art and still be unusable if it can’t survive the conditions it’ll actually face.

Edge AI is where cloud assumptions fall apart

Another session focuses on AI at the edge, where connectivity is limited and latency matters. That’s a polite way of saying cloud-native habits stop working.

Edge deployment forces a different architecture. You’re dealing with power, memory, thermal limits, and intermittent networking. In some settings, privacy and security requirements make remote inference a bad fit. That changes model choice, update cadence, monitoring, and rollout strategy.

Dr. Ali Agha, CEO and founder of FieldAI, is part of that conversation, along with Michelle Lee of Medra and Aidan Madigan-Curtis from Eclipse Ventures. The point isn’t the usual “edge is important” boilerplate. It’s the trade-offs.

On-device inference can cut latency and reduce dependence on the cloud, but it raises the cost of compression, quantization, and lifecycle management. Field updates are messier than pushing a new container to a cloud cluster. You don’t get to assume a stable network, a homogenous fleet, or easy observability.

If you’re building robotics, industrial AI, or aerospace systems, that’s not a footnote. It’s the product.

Prototype to production is where deep-tech startups get hit

The session on moving from prototype to production is probably the most useful one for founders and technical leads. It gets at the part too many demos skip: turning a working lab system into something that can be built, shipped, maintained, and sold at sane margins.

That’s where a lot of deep-tech startups stumble. The prototype proves feasibility. Production exposes yield, supply chain issues, tolerances, calibration drift, manufacturing cost, and support burden. The math changes fast.

The speakers here, including John Mackey of MBRYONICS, Boris Sofman of Bedrock Robotics, and Adrian Macneil of Foxglove, cover different slices of that pain. Space hardware, humanoid robotics, and autonomous systems all share the same problem: the first working version is rarely the expensive part. Making a thousand of them, or keeping them running in the field, is.

For engineers, that means thinking about:

• testability before feature count
• telemetry before rollout
• maintainability before novelty
• manufacturing constraints before roadmap fantasy

Plenty of startups can demo a robot. Fewer can ship one that doesn’t turn into a support headache.

Colossal brings the biology fight into AI

Then there’s the most controversial session on the stage: Colossal Biosciences CEO Ben Lamm on de-extinction, modern biology, and AI’s role in engineering nature’s comeback.

Colossal has become one of the more polarizing companies in tech, partly because the premise is so far out there and partly because the pitch sits between conservation, synthetic biology, and spectacle. That tension is exactly why it belongs here.

The technical question is how AI fits into the work. In biology, machine learning can help with protein design, genome analysis, phenotype prediction, and experimental prioritization. It can narrow the search space. It can’t erase biological uncertainty, ecological risk, or the ethical debate.

De-extinction also pulls attention away from conservation work that needs money now, not speculative moonshots. That doesn’t make the science dull. It just means the usefulness bar should stay high.

What this stage says about the industry

The Real World AI Stage is a reminder that “AI” has become too broad to be a single track. Software AI and physical AI share some tooling, but they split quickly once latency, failure modes, and the real world enter the picture.

That matters for senior engineers and technical leaders because the constraints are different.

If you’re building software AI, you care about model quality, data quality, security, cost, and product fit. If you’re building physical AI, add safety, certification, hardware integration, edge deployment, simulation, maintenance, and supply chain reality. That’s a bigger stack, and it’s harder to fake.

Disrupt is putting robots, defense, industrial AI, and de-extinction on the same stage because the interesting problems aren’t confined to chat interfaces and copilots anymore. The next wave runs on actuators, sensors, and systems that have to survive contact with the world.