Edge AI Is Starting to Move Beyond Pilots. NVIDIA Explains Where Scale Is Emerging First.

A robot that moves totes across a warehouse and a robot that folds laundry can both be described as physical AI. Commercially, they are years apart.

During NVIDIA's GTC Berlin press briefing, I asked Deepu Talla, the company's Vice President of Robotics and Edge AI, where NVIDIA sees the strongest shift from pilots to large-scale commercial deployments, which industries are leading, what is driving their return on investment and what still needs to change.

His answer was useful because it refused to treat robotics as a single market.

The constraint has not changed, he explained: "If the model is not sufficiently accurate and reasonably general purpose and safe, you can't really scale it up."

What has changed is that advances in end-to-end models, reasoning and simulation, refined through years of autonomous vehicle development, are beginning to make adjacent applications commercially viable.

The more interesting question is not whether robots are becoming intelligent. It is where that intelligence has become dependable enough to justify operating at scale.

Bounded tasks scale first

Talla distinguished several levels of maturity.

Autonomous navigation, mobile robots and inspection robots operating in unstructured environments are, in his words, "ripe for disruption." Manipulation of rigid objects, including pick-and-place operations, is also starting to gain traction.

General-purpose robots equipped with dexterous hands and expected to manipulate soft or unfamiliar materials remain a much harder challenge. Much of that work is still in research, even if, as Talla observed, "the best in the world are working on it."

Independent market data supports this distinction.

According to the International Federation of Robotics, almost 250,000 professional service robots were sold worldwide in 2025, an increase of 24%. Transportation and logistics accounted for approximately 117,500 units, representing 47% of that market.

By comparison, the IFR reported around 7,000 full-size humanoids sold for commercial and professional applications. Many current humanoid applications remain specialized or depend on human teleoperation.

These figures describe different robotics categories and should not be interpreted as a direct measure of AI autonomy. Nevertheless, they illustrate where commercial adoption is already established and where the technology remains experimental.

NVIDIA's own humanoid announcement reflects that distinction.

At GTC Berlin, Agile Robots introduced Agile ONE SN, a reference humanoid developed around NVIDIA Isaac GR00T and Jetson Thor. The platform combines a humanoid body, multimodal sensing and five-finger robotic hands, giving researchers an integrated environment for developing and validating new skills.

Its integration with NVIDIA's Halos safety system is planned, rather than already completed.

Scheduled for availability in early 2027, Agile ONE SN is intended to give universities, laboratories and smaller robotics developers access to capabilities that otherwise require substantial hardware and integration investment.

Talla acknowledged that developing an accurate, reliable general-purpose robot brain remains an unsolved problem.

The commercial opportunity is advancing, but it is advancing unevenly.

NVIDIA says its physical AI revenue reached nearly $11 billion during the first half of calendar 2026, more than 30% above the entire previous year. That broad revenue figure indicates momentum across NVIDIA's physical AI business. It does not measure the number of commercially deployed autonomous robots.

A more revealing indicator is whether the cost of developing and adapting autonomous systems is beginning to fall.

One example presented at GTC Berlin comes from ANYbotics.

Using NVIDIA Cosmos Data Factory, the industrial robotics company generated more than 1,500 synthetic video clips to help train inspection robots to recognize liquid spills and other changes between facility visits. The workflow uses Encord to curate real-world data and validate generated outputs.

The number of videos is less important than the problem being addressed.

Industrial robots need to recognize unusual conditions, but the examples required for training are often difficult, expensive or unsafe to reproduce. A facility cannot deliberately create every hazardous situation simply to improve an inspection model.

Synthetic data can expand the range of conditions represented during development, particularly when rare events are underrepresented in real-world datasets.

It does not eliminate the need for physical validation. It can, however, reduce one of the less visible barriers to scaling autonomy: obtaining sufficient representative data before a system is deployed.

The installed base is the market

Some of the most revealing numbers at GTC Berlin appeared during the industrial agents session.

NVIDIA presented an industrial landscape comprising an estimated 10 million factories, 100 million programmable logic controllers, 2 billion cameras and 5 million robots.

These figures cover different equipment categories and are NVIDIA's estimates, not directly comparable market measurements.

But they point toward an opportunity that receives less attention than humanoid robotics.

Much of physical AI's near-term value may come from adding intelligence to the cameras, controllers and production systems manufacturers already operate.

An existing industrial camera equipped with more capable perception and inspection software can potentially improve quality control without requiring an entirely new robotic installation. A production-planning agent can help optimize an established process before a company invests in new automated equipment.

Siemens' electronics factory in Erlangen, Germany, provides a particularly interesting example.

During the Physical AI session, Siemens presented the facility as a blueprint for AI-powered manufacturing.

The factory has operated since 1971. It handles high-mix, make-to-order, mid-volume production, conditions that make automation considerably more complex than in highly standardized mass manufacturing.

This is the reality facing many European manufacturers: decades of equipment, industrial controllers, software and operating procedures that artificial intelligence must work with rather than replace.

In its October 19 announcement with NVIDIA, Siemens described AI applications developed and tested in production at Erlangen.

For new printed circuit board assemblies, virtual production preparation could reduce engineering effort by up to 85% and time to market by up to 40%. Agent-based production planning could enable cost savings of up to 20%.

These figures are Siemens-reported, application-specific potential improvements. They should not be confused with independently verified returns across the manufacturing industry.

The underlying mechanism is more important than the percentages.

Preparing quality inspection traditionally required waiting for physical prototypes before collecting the images needed to train and configure inspection systems.

By using digital twins and NVIDIA Cosmos technologies to generate representative images earlier in the process, Siemens can prepare parts of that work before physical prototypes become available.

The same principle extends to production planning. An AI agent can generate alternative production scenarios, evaluate their potential consequences against defined indicators and use a factory digital twin to test proposals before human planners decide what to implement.

This brings attention to one of the least discussed components of industrial automation economics: the cost of change.

"Factories need to handle change as well as they handle repetition," observed Rev Lebaredian, NVIDIA's Vice President of Physical AI Simulation, in the Siemens announcement.

In a high-mix factory, automation costs are not limited to purchasing and operating equipment. Engineering teams must continually reconfigure processes, introduce new products, adapt inspection systems and manage unexpected conditions.

Reducing the engineering effort associated with those changes may deliver substantial economic value, even before more general-purpose robots become viable.

Talla made the European relevance explicit. Precision manufacturers and the Mittelstand of small and medium-sized industrial companies represent precisely the environments where, until recently, autonomous technology was not sufficiently mature.

Hybrid by design

The second lesson concerns where intelligence runs.

At GTC Berlin, Amit Goel, NVIDIA's Head of Robotics Ecosystem and Edge AI, presented new performance data illustrating the progress of local inference.

A chart showed selected open models running on Jetson Thor alongside more powerful cloud models on an intelligence index. Although the presentation did not provide enough methodological detail to establish equivalence across those models, it demonstrated how quickly embedded AI capabilities are advancing.

Another chart made the progress more concrete.

Against a reference line of 20 tokens per second for interactive generation, NVIDIA showed Jetson Thor delivering 84 tokens per second with Qwen 3.6 35B A3B and 133 tokens per second with Gemma 4 E2B in the configurations presented.

The mixture-of-experts architecture used by the Qwen model activates only part of its parameters during generation, helping explain why it can achieve higher throughput than some dense models in the same comparison.

These are NVIDIA-reported text-generation results, not independent benchmarks or measurements of robotic control-loop latency.

They nevertheless illustrate an important development: increasingly capable models can execute close to the machines using them, within defined memory, energy and performance constraints.

Talla was explicit that this does not mean everything will move onto the device.

Physical AI, he explained, will sometimes need to "phone a friend." That additional intelligence could come from another robot, an on-premises system within the factory, a telecommunications edge platform or the cloud.

"It's going to be a hybrid deployment."

The telecommunications announcements at GTC Berlin offer a concrete illustration.

Deutsche Telekom and Agile Robots demonstrated an architecture in which real-time perception and action remain on the robot, while distributed edge computing supports model fine-tuning and coordination across multiple robots. Centralized infrastructure supplies the computational capacity required for larger training workloads.

The demonstration does not establish commercial deployment at scale. Its significance is architectural.

It shows how robotic intelligence can be distributed according to the requirements of different tasks.

A robot cannot depend on unpredictable network delays for collision avoidance or other time-critical safety functions. Those functions must remain local or be handled by suitably designed independent safety systems.

But coordinating a fleet, adapting models or processing large amounts of industrial data may benefit from resources beyond the individual device.

A factory server could coordinate several mobile robots. A telecommunications edge site could support distributed industrial applications. Centralized infrastructure could run computationally intensive training or simulation without becoming essential to the safe operation of every machine.

This is also changing the role of telecommunications operators.

NVIDIA is working with operators across EMEA to develop distributed AI infrastructure, extending accelerated computing from central data centers toward regional and network-edge locations.

The maturity of those initiatives varies. Elisa has conducted an over-the-air AI-RAN field trial combining communications and AI workloads. Deutsche Telekom has validated related technology in laboratory environments, with field trials planned for 2027. BT also plans live trials in 2027.

These should not be confused with widespread commercial deployments.

For European industrial customers, however, the direction is strategically relevant.

The ability to place computing resources within a factory, regional network or sovereign infrastructure may influence latency, resilience, operating cost, data governance and control over proprietary industrial information.

The engineering principle is to locate each function according to what it must guarantee.

Safety-critical decisions must remain dependable during network interruptions. Higher-level planning and optimization can use distributed resources, provided the system degrades safely when connectivity or remote services become unavailable.

A hybrid architecture may improve capability and flexibility, but it also introduces dependencies across infrastructure, cybersecurity and operational responsibility.

Those dependencies must be engineered, not assumed away.

Reliability you can prove

The third lesson will determine how quickly the market develops.

Consider two fleets of mobile robots performing 5,000 assignments per day.

One requires human intervention on 10% of assignments. The other requires it on only 1%.

If each intervention takes three minutes, the first fleet generates 25 operator-hours of intervention daily. The second generates two and a half.

This is a hypothetical illustration, not a reported deployment result. But it shows why seemingly modest improvements in reliability can dramatically affect operational economics.

The cost of a successful autonomous task includes much more than the price of the robot. Integration, energy consumption, maintenance, exceptions, software updates, downtime and safety validation all influence the business case.

Goel's presentation on NVIDIA's Halos safety architecture also demonstrated how many layers are required to achieve dependable physical AI.

The architecture combines safety-related hardware, operating-system mechanisms, independent control functions and evaluation methods for learned models.

NVIDIA presented specific safety-related development and hardware capabilities associated with IEC 61508, while acknowledging that some elements of its robotics safety software stack were not yet available.

These capabilities concern particular components, development processes and fault categories. They do not constitute certification of an entire robot.

The architectural implication is important.

A learned model can interpret complex surroundings and propose actions, but safety mechanisms must remain capable of constraining or overriding that behavior when necessary.

This is why another GTC Berlin announcement deserves attention.

NVIDIA introduced Halos SAFE, the Safety Assurance Framework for Evaluation, extending its Halos safety approach with reusable evaluation workflows for autonomous vehicles.

Developed with collaborators including dSPACE, IVEX and TÜV Rheinland, Halos SAFE connects simulation, software-in-the-loop and hardware-in-the-loop testing, on-road evaluation and fleet monitoring.

The objective is to use defined scenarios, measurable criteria and traceable results to build stronger evidence of system behavior.

In my earlier IEEE Computer article on NVIDIA Halos, I examined the broader question of how safety can be engineered across the complete physical AI stack.

What makes Halos SAFE relevant now is its attempt to establish a more explicit and reusable evaluation framework.

An open evaluation methodology does not automatically mean that a system has been independently certified. Nor does extensive simulation eliminate the need to validate behavior in real operating environments.

But stronger evaluation methods can help manufacturers, assessors and customers understand the evidence supporting a safety claim.

Talla indicated that NVIDIA intends to extend related approaches to general-purpose robotics.

That could become a significant enabler of adoption.

A manufacturer cannot build a compelling business case for autonomous equipment solely on the promise that its AI performs well in typical conditions. It also needs to understand what happens when the system encounters uncertainty, fails to recognize an object, loses connectivity or must transfer control to a human operator.

Those are engineering requirements, but they also determine commercial risk.

What executives should take from Berlin

Three questions emerge from NVIDIA GTC Berlin.

Is the task sufficiently bounded that success and failure can be specified and measured? Is the architecture designed to distribute intelligence without compromising safety? And can reliability be demonstrated with evidence that an independent party could scrutinize?

These questions offer a more useful framework for evaluating physical AI investments than the sophistication of an individual demonstration.

General-purpose humanoids will continue expanding the boundaries of robotics, and research platforms such as Agile ONE SN may help accelerate that progress.

The industrial opportunity available today, however, lies in operations where autonomy can already be specified, integrated and validated.

The next competitive advantage will not necessarily belong to the manufacturer deploying the most advanced robot. It may belong to the company that makes existing equipment more intelligent, reduces the cost of adapting production and combines local autonomy with reliable industrial infrastructure.

The milestone to watch is the moment a task that impressed on stage starts running every shift, at a cost a plant manager is willing to approve.

Related analysis by Stéphane Gervais

·        Building Trust Into AI Systems: Inside NVIDIA's Halos for Physical AI Safety, IEEE Computer, 2026.

·        Why Physical AI Forces Executives To Rethink Safety, Accountability And Trust, Forbes, February 2026.

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

·        En 2026, l'IA sort des écrans pour transformer le monde physique, Alliancy, January 2026.

Disclosure: Stéphane Gervais is VP Strategic Partnerships at CertX, an SGS Group company. CertX participates in the NVIDIA Halos AI Systems Inspection Lab ecosystem. This article reflects the author's independent analysis.

Reporting: NVIDIA GTC Berlin news pre-brief and Physical AI sessions, October 20, 2026. Additional sources: NVIDIA GTC Berlin announcement drafts, International Federation of Robotics and Siemens.