Physical AI is artificial intelligence that doesn’t just describe the physical world but acts on it: perceiving conditions, reasoning about response options, and executing physical change. NVIDIA’s broader stack (Omniverse and Cosmos for simulation, DRIVE AGX Thor for embodiment, DGX for refinement) has made the compute foundation possible, and the Alpamayo Open Platform (the Alpamayo models, the AlpaSim simulation framework, the AlpaGym training framework, and the Physical AI Dataset), with the Cosmos-Dreams world model, now turns that compute into AV-specific tooling. As Jensen Huang put it at GTC Berlin in October 2026, simulation is “where we have the AI learn how to do its skills,” inside a digital twin, before it acts in the physical world. The skills vehicles will perform with that compute remain to be defined.

AV Exit’s IP defines transformation behavior: skills no vehicle’s AI has learned yet, because no vehicle has performed them yet. Not as software running on top of NVIDIA’s stack, but as patented and patent-pending architectural definitions of what a self-reconfiguring vehicle can do.

FIVE SCENES FOR FIVE BEHAVIORS

Picture an autonomous vehicle pulling into a driveway, detaching its cabin onto temporary supports, accepting a different cabin configuration, and driving away. Without human intervention. No depot, no fixtures. On a flat surface that didn’t need to be prepared for the operation. That’s Swap by AI™.

Picture a vehicle whose passenger doors open and close not via dedicated motors but via the same wheels that drive it down the road, eliminating an entire mechanical subsystem and its lifetime failure modes. That’s Enter by AI™.

Picture a vehicle that perceives an imminent, unavoidable impact and, before contact, extends AI-controlled bumpers, moves its cabin back on the chassis, and meters the crash energy segment by segment, partially separating the cabin if needed to keep the force away from its occupants. That’s Safe by AI™.

Picture a parked vehicle with a solar-energy collecting photovoltaic roof, whose cabin detaches and tilts onto support legs, revealing additional photovoltaic surface on the platform itself, while the cabin remains usable for working, sleeping, or storage. That’s Charge by AI™.

Picture a vehicle whose AI sees an oncoming car it cannot avoid and, with seconds to spare, sends the entire passenger pod rearward off the chassis, leaving the empty platform to take the impact. That’s Exit by AI™.

WHERE THESE SKILLS ARE LEARNED

NVIDIA’s Alpamayo, the autonomous-vehicle reasoning model announced by Jensen Huang at CES 2026 in January, is the first NVIDIA full-stack AV deployment to reach a production vehicle: the all-new Mercedes-Benz CLA, debuting NVIDIA’s full-stack driving software (then called DRIVE AV, now marketed as the Alpamayo Stack) with a U.S. Level 2++ launch in 2026. At GTC Berlin in October, Mercedes-Benz announced it is extending the partnership to Level 3 and Level 4 development.

At GTC 2026 Taipei in June, the model family advanced to Alpamayo 2 Super (an open 34-billion-parameter reasoning model with full 360-degree perception, available on Hugging Face). At GTC Berlin in October, NVIDIA added Alpamayo 2 Edge, a compact reasoning model that runs on the vehicle’s own DRIVE AGX Thor, and automakers including Volvo Cars and Mercedes-Benz adopted the Alpamayo Stack, NVIDIA’s full-stack driving software on DRIVE Hyperion. Alpamayo demonstrates that AVs can be trained to reason from a few million simulation miles rather than billions of real-world miles.

NVIDIA’s AlpaSim, the open-source closed-loop simulation framework released at CES 2026 and available on GitHub, orchestrates the environments in which Alpamayo’s policy decisions get tested in real time.

NVIDIA’s Cosmos-Dreams, real-time generative closed-loop AV simulation built on Cosmos, generates the photorealistic, action-conditioned views inside the simulation loop. Text-prompt scenario authoring lets developers describe what should happen and then watch the simulated “ego” AV respond.

NVIDIA’s AlpaGym, the closed-loop reinforcement-learning framework introduced at GTC 2026 Taipei in June, post-trains the policy through continuous decision-and-observation cycles inside AlpaSim.

NVIDIA’s Physical AI Dataset, another component of the Alpamayo open platform, supplies the real-world driving data Alpamayo learns from: 300,000 clips spanning 1,700 hours across 2,500 cities in 25 countries. The Dataset captures what fleets do today (lane following, merging, braking, urban navigation). Transformation behaviors lie beyond what any fleet performs yet.

NVIDIA Halos, NVIDIA’s full-stack safety system for autonomous vehicles, gained a new framework at GTC Berlin, now part of the Alpamayo Stack: Halos SAFE (Safety Assurance Framework for Evaluation), which links simulation, hardware-in-the-loop testing, on-road testing, and fleet monitoring to turn defined scenarios, tests, and measurable performance criteria into traceable safety evidence, with evaluation partners including dSPACE, IVEX, and TÜV Rheinland. Safe by AI™ and Exit by AI™ are defined in exactly those terms: specific crash conditions, specific responses, measurable outcomes.

AV Exit’s IP is the missing content for this open platform. These behaviors cannot be learned from observation, but they can be defined architecturally by AV Exit’s IP, run as closed-loop scenarios in AlpaSim, rendered photo-realistically by Cosmos-Dreams, post-trained through AlpaGym, evaluated under Halos, and used to train Alpamayo’s reasoning. Once Alpamayo learns the behaviors, DRIVE AGX Thor onboard compute executes them on modular vehicle hardware in the real world.

WHY NOW

The EV transition has been disrupted by policy shifts, tariff regimes, charging infrastructure gaps, raw materials volatility, demand shortfalls, and write-downs across multiple legacy automakers. The industry pattern is now visible: incremental EV strategies face structural headwinds that no single OEM can solve on its own timeline. The strategic question for an OEM facing that pattern is not how to fix EV-1.0. It’s how to leapfrog past it into AV Level 4 (hands-off, eyes off), and beyond.

Physical AI is the leapfrog. It is built to move. Modular autonomous vehicles built on NVIDIA’s compute stack and AV Exit’s transformation behavior IP represent a new paradigm, self-reconfigurable modularity: vehicles whose architectural moves create capabilities no unibody vehicle can mechanically perform. An OEM committing to this approach isn’t catching up on EVs. It’s defining the next category while competitors are still solving the last one. NVIDIA supplies the platform. AV Exit supplies the paradigm.

FROM PARADIGM TO PRODUCTION

NVIDIA’s DRIVE Hyperion has become the common platform for the autonomous age. At GTC Taipei in June 2026, Jensen Huang said car manufacturers representing roughly 80% of the world’s vehicles have adopted DRIVE Hyperion, NVIDIA’s Level 4-ready platform. These are the OEMs that would build modular AVs, and the manufacturing base that would build Swap by AI vehicles. BYD, Geely, Nissan, and Hyundai (four of the world’s top ten automakers) adopted DRIVE Hyperion for Level 4 development in March. In October, Volvo Cars expanded its NVIDIA partnership, and at GTC Berlin NVIDIA announced that Volkswagen, JLR, Volvo, and Mercedes-Benz vehicles will join its Hyperion robotaxi network, with Bolt planning at least 25,000 Level 4-capable Lucid vehicles in Europe. As Jensen Huang said in Berlin, “every single car company will want to make their cars part of the ride hailing fleet.” A ride-hailing fleet is where a swappable pod earns its keep: one platform, a passenger pod by day and a cargo pod by night.

The NVIDIA-Hyundai partnership announced in October 2025 anchors South Korea’s National Physical AI Cluster, a public-private partnership deploying roughly $3 billion and 50,000 NVIDIA Blackwell GPUs. The partnership encompasses Kia, whose Platform Beyond Vehicle line builds passenger, cargo, and chassis-cab versions on one platform, with each body fixed at the factory or by a conversion partner. Kia’s Easy Swap concept would exchange rear modules at a purpose-built facility. Swap by AI™ enables autonomous pod exchange anywhere, no depot required. Enhanced consumer convenience. Reduced OEM infrastructure cost.