The Platform War Behind Robotics: How NVIDIA Plans to Become the Infrastructure of Physical AI

The most consequential competition in robotics may not be between robot manufacturers. It may be over the infrastructure that turns physical experience into scalable intelligence: data, simulation, models, edge compute, safety and standards.

Robotics is usually presented as a race between machines: humanoids versus industrial robots, one robot maker against another, one breakthrough in dexterity or autonomy against the next. That framing increasingly misses the deeper strategic contest.

The global installed base of industrial robots reached 4.664 million units in 2024, according to the International Federation of Robotics, with another 542,000 installed during the year. Much of that installed base remains optimized for narrow, repeatable tasks rather than general-purpose adaptation.

Physical AI changes the economics of that model. As robots gain broader reasoning and adaptation capabilities, the complexity shifts from programming individual movements toward building and maintaining the infrastructure through which machines learn: collecting data, training models, simulating environments, evaluating policies, deploying intelligence at the edge, improving performance and proving that the resulting system is safe. That creates a different kind of market.

NVIDIA is positioning itself not simply to sell compute into that market, but to become one of its foundational technology layers. My conversation with Sasa Docca, from NVIDIA's robotics product marketing team, at MACHINA 2026 made the strategy unusually explicit.

"We're an accelerated platform computing company."

The important word is not computing. It is platform.

Robotics is moving from vertical integration to a platform economy

Historically, building an advanced robot required companies to solve an extraordinary number of problems themselves. Beyond mechanics, actuators and sensors, they needed perception, control, compute, data pipelines, simulation, training infrastructure, deployment software and increasingly sophisticated safety systems. That degree of vertical integration was partly unavoidable because reusable infrastructure was immature or fragmented.

NVIDIA's strategic answer is not to force every robotics company onto one monolithic stack. Docca described the opposite approach:

"We're going to disaggregate this whole platform. So then you can work with us at any level."

A company building a robot brain may primarily need accelerated compute. An OEM may use NVIDIA simulation frameworks while retaining its own control stack. Another company may adopt NVIDIA Jetson for onboard inference, NVIDIA Cosmos for data generation or NVIDIA Halos components for safety. The point is modularity: every layer becomes a potential entry point without requiring the customer to surrender the rest of its architecture.

The breadth of the ecosystem already illustrates that ambition. NVIDIA says companies spanning industrial robotics, humanoids, surgical systems and robot foundation models are building with its technology, including ABB, FANUC, KUKA, Yaskawa, Agility, Figure and Skild AI. ABB, FANUC, KUKA and Yaskawa alone have a combined installed base exceeding two million industrial robots.

This is why the platform opportunity is much larger than humanoids. NVIDIA does not have to predict which robot form factor ultimately captures the most value. It has to make its infrastructure useful across them.

The CUDA playbook, but with a major difference

There is an obvious analogy with NVIDIA's position in digital AI. CUDA became strategically important not because NVIDIA built the applications running on GPUs, but because it provided a programming and acceleration layer on which others could build. Robotics presents a similar opportunity: absorb complexity below the application, provide reusable infrastructure and let many companies compete above it.

But physical AI is fundamentally harder. There is no single equivalent of a language token for the physical world. Robots differ in morphology, sensors, actuators, degrees of freedom, environments and safety constraints. Training data is harder to acquire. A software failure can become a physical event. A policy that works in simulation can still fail when confronted with friction, deformation, contact, occlusion or an unexpected human action.

The robotics platform therefore has to extend much further than compute. It must connect the digital development loop to physical execution. That is why NVIDIA's robotics portfolio now spans data capture, open robot models, physics-based simulation, world models, edge inference, deployment frameworks and safety infrastructure.

The platform war is therefore not a conventional product contest. It is a contest over which infrastructure becomes the easiest place to move from physical experience to reusable intelligence, and then back into machines operating in the real world.

Openness is becoming part of the competitive strategy

Perhaps the most interesting part of NVIDIA's approach is its emphasis on open robotics infrastructure. Docca was explicit:

"The next phase is going to be fueled by open models and open frameworks. That's what you need."

This is not simply an open-source philosophy. It is market architecture. In July, NVIDIA and Hugging Face made NVIDIA Isaac Teleop and Isaac GR00T 1.7 available through LeRobot, Hugging Face's open-source robotics framework. GR00T 1.7 is an open, commercially licensed reasoning vision-language-action model, while Isaac Teleop is an open framework for collecting human demonstrations that can be used to post-train robot policies.

NVIDIA has also announced an open Isaac GR00T reference humanoid that combines a Unitree H2 Plus platform, Sharpa Wave tactile hands, NVIDIA Jetson Thor compute and GR00T software and workflows. The strategic point is not that NVIDIA becomes a humanoid OEM. It is that researchers can work from a more consistent reference architecture rather than rebuilding every hardware-software interface from scratch.

The same pattern is visible in simulation. Newton, an open-source GPU-accelerated physics engine co-developed by NVIDIA, Google DeepMind and Disney Research and managed by the Linux Foundation, is compatible with NVIDIA Isaac Lab and Isaac Sim. NVIDIA can participate in an open ecosystem without needing to own every component of it.

Strategically, openness solves several problems at once. It expands the developer base, lowers integration friction and makes datasets, models and workflows easier to reuse across embodiments. It also creates a deliberate trade-off: open interfaces make components easier to adopt, but also easier to replace. In a fragmented industrial market, NVIDIA appears to be betting that becoming the most useful substrate is more valuable than forcing full-stack lock-in.

That can still produce strong platform economics. If open tools increase the number of developers training, simulating and deploying physical AI, they increase demand for compute, optimized runtimes and edge systems across the lifecycle. The interface can be open while the underlying infrastructure remains economically valuable.

The real scarcity is increasingly data

The transition from specialist robots toward what Docca calls "generalist specialists" is central to this strategy. Today, many industrial robots are exceptionally good at one task. The longer-term objective is broader intelligence that can still reach industrial levels of performance in specific applications. Getting there requires data on a scale and diversity that conventional robotics has never needed.

Language models benefited from vast quantities of existing digital information. Robotics has no equivalent corpus of physical experience waiting to be downloaded from the internet. A robot needs demonstrations of actions, observations from its own sensors, information about physical interactions, failures, edge cases and the consequences of different decisions that reflect the scale and diversity of the real world.

That is why the robotics data pipeline is becoming strategically important. As Docca put it:

"Teleop is one way of collecting real data, but then the world models help you scale that."

Teleoperation can capture valuable human demonstrations. Physics-based simulation can create controlled, repeatable training environments. World models can expand coverage by generating variations, plausible future states and situations that would be costly, rare or unsafe to reproduce physically.

NVIDIA's Cosmos 3 is a clear expression of that direction. The open world foundation model combines vision reasoning, world generation and action prediction in one architecture. In July, NVIDIA added Cosmos 3 Edge, a smaller model designed to bring on-device vision reasoning and robot policy capabilities to Jetson-class edge systems.

But world models are not a replacement for physics. In the interview, Docca emphasized that world models themselves are grounded in real data and classical simulation. The emerging physical-AI stack will therefore draw on multiple sources of learning: real demonstrations, digital twins, physics simulation, synthetic data, world models and eventually operational fleet experience. The platform that can connect those sources has leverage.

Asked what would separate robotics companies that scale into industrial operations from those stuck in pilots, Docca came back to the same issue: a strong data strategy and the ability to turn compute into data. That is a useful way to frame the next competitive bottleneck. Compute alone is not the scarce asset; the question is how effectively it can be converted into task-relevant experience.

The next battle is over interfaces and standards

Platforms become powerful when different parts of an ecosystem can interoperate. That is especially important in robotics because fragmentation remains severe: data formats differ, robot bodies differ, CAD and simulation environments differ, deployment software differs and safety approaches differ.

Docca sees standardization as one mechanism for reducing that fragmentation:

"Standardization helps because then more people can jump on it."

He pointed to several layers where common interfaces and de facto conventions are emerging: the LeRobot dataset format for robot learning, OpenUSD for exchanging complex 3D assets, and ROS 2 as a widely adopted open-source robotics framework. These are not the same kind of standard, and none is a universal solution, but each reduces the engineering friction between layers.

That has strategic consequences. Interfaces determine how easily data moves between tools, how readily one component can be replaced by another and how much integration work is required to connect a stack. For NVIDIA, interoperability makes its technology easier to insert into heterogeneous robotics architectures. For OEMs, the same interoperability preserves optionality and reduces dependence on any single supplier.

A successful platform must therefore balance two forces that are often in tension: enough integration to create a compelling developer experience, and enough openness to let robot makers preserve the layers where they differentiate. NVIDIA's disaggregated approach is designed around that balance.

Safety may become a platform layer too

Safety is where the analogy with conventional AI infrastructure breaks down most sharply. A generative model can produce an incorrect answer. A physical AI system can move a machine. For robots operating around people, safety extends across compute, sensors, operating software, applications, AI behavior, runtime monitoring and the environment itself.

NVIDIA launched Halos for Robotics in June 2026 as a full-stack functional-safety system for physical AI. The architecture combines IGX Thor safety-capable compute, Holoscan Sensor Bridge, Halos OS and safety application blueprints, together with the Halos AI Systems Inspection Lab.

The architecture treats safety in two directions. Inside-out safety addresses the robot itself: its ability to perceive risks, respect constraints and transition to a safe state. Outside-in safety extends awareness beyond onboard sensors by using external cameras and infrastructure to detect hazards that may be occluded from the robot.

NVIDIA says Agility's Digit will be the first production robot shipping with Halos OS. The Halos AI Systems Inspection Lab is an ANAB-accredited ISO/IEC 17020 Inspection Body, providing a structured inspection path before independent third-party system certification. That distinction matters: Halos can provide reusable, preassessed safety building blocks, but it does not remove the OEM's responsibility for the safety case of the complete robot in its intended environment.

The strategic implication is larger than Halos itself. If physical AI is to scale across factories, warehouses, hospitals and eventually public environments, safety assurance cannot remain a fully bespoke process recreated independently for every machine. Reusable architectures, inspection methods, validation workflows and certification pathways could become as important to market scaling as common software frameworks.

That is the infrastructure side of the story. The operational side, how data, safety and platform tooling translate into uptime, service models and ROI, is examined separately in The Next Humanoid Robotics Race: Safer Robots, Bigger Scale, More Markets.

Agentic AI could become the orchestration layer

There is another layer beginning to emerge above this stack: agents. During the interview, Docca described a development process in which an agent can coordinate parts of the robotics workflow without engineers manually orchestrating every tool. His example was deliberately simple:

"When I speak to an agent, I just say, 'Spin up Isaac Sim for me.'"

The deeper idea is more consequential. A deployed robot encounters a task it cannot complete. That experience can be captured. An agent can help analyze the failure, assemble relevant data, create new scenarios in simulation, initiate evaluation and support a retraining workflow. Docca summarized the shift this way:

"I no longer have to be very prescriptive about the workflow."

The idea is already moving from concept into tooling. NVIDIA JetPack 7.2 made Jetson "agentic-ready" for robotics, inspection and industrial automation. NVIDIA's current Halos Outside-In Safety documentation also includes an agentic deployment path in which compatible coding agents can bring up defined safety profiles from natural-language instructions.

That should not be confused with allowing a software agent to modify a production robot without control. In safety-critical physical systems, retraining, validation, change approval and deployment gates remain essential. The more important possibility is that agents become connective tissue across the engineering workflow, spanning data collection, simulation, optimization, testing and edge deployment.

If that happens, the value of an integrated platform increases again because the orchestration layer needs standardized ways to invoke each underlying capability.

What NVIDIA cannot turn into a platform

None of this means robotics becomes a commodity. The robot body still matters. Actuation matters. Manipulation matters. Mechanical design matters. Low-level control matters. Application expertise and proprietary operating data matter. So do integration, maintenance, service operations and the ability to solve a customer workflow reliably.

Docca was clear that NVIDIA does not expect partners to give up those layers. In his words, robot makers "have their own pipelines" and "their own vision of what they want to do." The role of the platform is to fill specific gaps, not to erase the OEM.

This boundary matters because robotics is unlikely to reproduce digital AI exactly. Physical systems remain heterogeneous. OEMs have strong incentives to protect the layers where their differentiation resides. Industrial customers often need multiple suppliers. Open frameworks make substitution possible as well as adoption. Safety responsibility ultimately belongs to complete systems operating in defined environments, not simply to the compute platform underneath them.

NVIDIA's success is therefore not guaranteed by the breadth of its portfolio. Its challenge is to make each layer valuable enough to adopt while keeping the overall architecture open enough for robotics companies to preserve their independence. Paradoxically, the disaggregation Docca describes may be precisely what makes that possible.

Who captures the value, and what that means for the rest of the industry

This leads to the central strategic question for the robotics industry: as physical AI matures, where will the durable economic value accumulate?

Some will remain with robot manufacturers that master hardware and industrialization. Some will accrue to companies with proprietary robot intelligence or valuable domain data. Integrators and operators will capture value from applications, fleet operations and workflow transformation. Infrastructure providers occupy a particularly attractive position because they can benefit from growth across competing robot categories.

NVIDIA does not need a single humanoid manufacturer to dominate. It needs increasing numbers of robots, of many kinds, to require accelerated training, simulation, synthetic data, world models, edge inference and safety infrastructure. That is the core logic of the platform bet.

The "war" is therefore being fought across several control points. Chip vendors want the compute layer. Cloud providers want training and fleet workloads. Foundation-model developers want the intelligence layer. Automation incumbents want the industrial workflow. Robot OEMs want to protect embodiment, control and operating data. Integrators want to own the customer problem. NVIDIA is trying to connect enough of these layers that it participates regardless of which robot category wins.

For robot OEMs, the strategic question is increasingly which layers to keep proprietary and which to source externally, since every layer built in-house must be maintained and continuously justified against increasingly capable shared infrastructure. For enterprise buyers, the platform beneath a robot is becoming material to due diligence because it affects portability, update cycles, safety validation, data ownership and the ability to adopt new models without costly re-architecture.

For NVIDIA's competitors, silicon alone is unlikely to be enough. A credible alternative will also need a development toolchain, simulation and data workflows, edge deployment, interoperability and a path to safety assurance. At the same time, the openness that expands NVIDIA's reach also gives the ecosystem alternatives. The outcome is not predetermined.

That is why NVIDIA's strategy should be viewed as more than a robotics product portfolio. It is an attempt to shape the architecture around which a new computing market develops.

The real prize is the architecture beneath the robot

The visible robotics race will keep producing dramatic machines: more dexterous humanoids, more adaptive industrial arms, more intelligent autonomous mobile systems and forms that do not yet exist. But the most durable competitive position may belong somewhere less visible than any single robot on a factory floor.

In every major computing transition, significant value has accrued not only to the companies building final products, but also to those that created reusable infrastructure on which many products depended. Physical AI is creating an opportunity for a similar layer. NVIDIA is assembling it across accelerated compute, open robot models, teleoperation, simulation, world models, edge deployment, safety and increasingly agentic orchestration.

Its decision to make much of that architecture modular and open is not peripheral to the strategy. It is what allows the platform to enter an industry that will remain heterogeneous by nature. But robotics is more fragmented than digital AI, physical systems impose harder constraints than software, and OEMs will continue to defend the technologies and data that make their products unique.

The next robotics battle will not be decided only by who builds the best machine. It will also be decided by who controls the critical interfaces connecting data, models, simulation, compute, safety and deployment.

The strategic prize is no longer simply the robot. It is becoming the architecture beneath it.