Nvidia's edge and physical AI bonanza

Edge computing and IoT aren't new, but what is new is the availability of more powerful, more compact GPU processors and small language models suitable for remote regions of the network, from cell phone network towers to mobile devices to sensors on factory-floor robots.

Among the many big splashes for Nvidia at its GTC conference in March was the release of its RTX Pro 4500 chip, designed for light data center inferencing and edge inferencing, with half the power draw of the higher-end RTX 6000 released in 2025. The RTX Pro 4500 chip was unveiled with pledges of support from most of the major enterprise IT vendors, including Cisco, Dell and HPE.

"As you get deeper in the network, we want to be able to put potentially more cost-effective, power-effective GPUs into servers," said Kevin Wollenweber, senior vice president and general manager of data center and internet infrastructure at Cisco, during an interview with Informa TechTarget in March.

Late last year, Cisco launched a set of unified edge devices that have compute, storage and networking built in for more remote deployments, and the RTX Pro 4500 will be slotted into some of those devices, Wollenweber said.

"For some of these localized use cases, video transcoding or video analysis, we can actually put GPUs inside of a warehouse or in remote sites to do localized processing using AI with smaller language models, versus pulling those back to a data center with LLMs," he said.

Edge AI's implications for enterprise

Enterprises already have their hands full scaling AI in data centers, and few have run AI inference at scale in large edge and IoT environments so far, but edge AI still became the subject of growing buzz in the industry this year following GTC.

"At KubeCon in Amsterdam, we were talking a lot about cloud factories reaching edge locations reliably and economically," said Torsten Volk, an analyst at Omdia, a division of Informa TechTarget. "That opens up a whole range of AI use cases that need low-latency responses to actually work, [such as] visual inspection of items during a production process, direct customer interaction in retail, controlling medical equipment in hospitals, etc."

Once organizations start running inference locally and see what's possible, the use cases multiply way beyond what anyone planned for initially.

Systems developed for AI inference efficiency also potentially have a unique appeal to enterprises with less rigorous demands -- and smaller budgets -- than frontier model developers, said Mike Leone, an analyst at Moor Insights & Strategy, who was an analyst at Omdia at the time of his March interview with Informa TechTarget.

"Most enterprises are never going to build massive GPU clusters, and they don't need to," he said. "The RTX Pro 4500 changes the equation because it puts cost-effective AI inference into standard workstations and edge servers without specialized cooling or a facility retrofit. Think manufacturing floors, hospital imaging labs, retail back offices. Once organizations start running inference locally and see what's possible, the use cases multiply way beyond what anyone planned for initially."

Physical AI edges toward reality

During the first quarter of 2026, Cisco, Dell, HPE and Nvidia also launched updates for AI inference in radio area networks (RAN), including partnerships with telcos such as AT&T and T-Mobile to push more processing power into edge networks, with the goal of enabling physical AI.

We're starting to see early but meaningful signals, especially in manufacturing, robotics and environments where decisions directly interact with the physical world.

Physical AI integrates with systems such as robots, drones and autonomous vehicles, using sensors and generative AI models to perform autonomous operations. Alongside hardware updates at GTC, Nvidia showcased its Omniverse collection of libraries, blueprints and microservices for developing physical AI applications . Omniverse includes tools such as the open source Isaac AI robot development program and the Newton Physics Engine, along with Cosmos world models and Groot robotics foundation models.

This is an even more nascent area than edge AI for enterprises, where significant reliability challenges remain, said Varun Raj, head of private cloud platforms, AI platforms, infrastructure and enterprise workloads for a global consulting firm.

"We're starting to see early but meaningful signals, especially in manufacturing, robotics and environments where decisions directly interact with the physical world," Raj said. "What's notable is that these systems amplify the same failure modes, but with far less tolerance for correction. That's likely to drive a convergence between AI systems and traditional control systems engineering, where bounded behavior and predictable failure modes matter more than raw capability."

In the meantime, companies such as Booz Allen Hamilton are already working with telcos on 6G cellular technology, which will combine with edge and physical AI to transform mobile computing within the next three to five years, with potentially industry-shifting results, according to Bill Vass, CTO at Booz Allen.

"In some ways, AI RAN is a component of 6G, and it is transformational for the way the cell phone towers will work in the future," Vass said. "Trees affect them, buildings affect them. … By putting AI at the edge, the cell phone tower can effectively learn its environment by measuring reflections and attenuation to its users, and it can increase its coverage – it's an important part of the next generation of cell service."

6G, AI and the next generation of the internet

With more computing power in cell phone towers, the processing power of edge devices, from mobile phones to robots, will not be limited by local hardware resources, raising heady possibilities for the future of software and applications from consumers to enterprises, Vass said.

"You can imagine adding all sorts of different types of sensors to the cell phone tower to let it do more than just AI RAN and cellular communications and processing, now that you've got GPUs there," he said. "If you're deploying a remote application as part of your business for your customers, you might be able to say, 'My customers' cell phones don't have enough GPUs to run this entrance app, but I can run it on the tower,' and it'll feel just like it's on their phone because it's so low latency between the phone and the tower."

Other companies are already developing such applications, such as Personal AI, which is working with Comcast's AI Grid, first targeting consumers and small businesses, but ultimately enterprises as well, according to its CEO, Suman Kanuganti.

"For example, a small business owner who has three or four phone lines, who doesn't have enough staff to pick up 80% of calls, has $100,000 lost income on a yearly basis," Kanuganti said. "If you put AI in their existing phone, that is private to them, with data that is specific to them, about their prices, their listings, that grows and evolves with them and also establishes the [customer] relationships."

Larger companies could create similar personalized AI for employees and use physical AI to boost what virtual workers can do, Kanuganti said. Edge AI delivered via telecom networks could also avoid some of the cybersecurity headaches enterprises have encountered with AI in data centers.

The AI grid becomes … a necessity for providing this next generation of internet in the form of AI, democratized to everybody.

"Telecom is a highly regulated industry -- every phone call that goes in is governed by federal regulations, so there are zero chances for that data to be going to any cloud whatsoever," he said.

Moreover, telecom networks have already built the infrastructure to absorb AI workloads, according to Kanuganti.

"Telcos also already have distributed power. They already have distributed real estate. They have distributed compute, and more importantly, they have the data layer, the pipes that actually move the data from place to place," he said. "[As] we start creating the next generation of workflows, we anticipate going from whatever internet traffic today is to five times that with AI. That's where the AI grid becomes not a fancy toy, as much as it becomes a necessity for providing this next generation of internet in the form of AI, democratized to everybody."

From there, the implications for individuals and enterprises alike are, in some ways, still unimaginable right now, according to Vass, who has previously held leadership roles at AWS and the Department of Defense.

"When I was at the Pentagon, we commercialized GPS, and I remember telling people that everyone will have a laptop screen in their car," Vass said. "They thought I was crazy, but I never thought about Lyft and Uber and DoorDash -- it's very hard to know, when you enable a capability, how people will use it."

Beth Pariseau, senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism. Have a tip? Email her or connect on LinkedIn.