NVIDIA GTC AI CONFERENCE, SAN JOSE, Calif. — A cute robot almost stole the spotlight away from Nvidia's CEO Jensen Huang during today's keynote for the company's big AI developers' conference. "Almost" is the key word here because we all know that Huang has quickly become a tech superstar — and the packed SAP Center stadium in San Jose showed him the love, especially when he took the stage and gleefully fired a T-shirt cannon into the audience of thousands before espousing the transformational nature of AI combined with communications.
"AI can revolutionize communications," he said. "There is no question in my mind that accelerated computing infused with AI will do a far better job adapting radio signals and massive MIMOs to changing environments and traffic conditions."
Because the emerging generation of robots and other AI applications will require massive amounts of data, Nvidia is working with telcos and networking vendors to transform networks to meet those requirements — using AI to serve AI.
Nvidia is working with 150 telcos around the world, including 90% of the top 50, and they are rapidly adopting AI across their business for internal productivity, customer experience and improving performance and performance-per-watt on the wireless network, Nvidia SVP Ronnie Vasishta said in a briefing for press and analysts Monday.
"The countdown to 6G has already begun," Vasishta said. "Fundamental research has already turned its attention to wireless telecommunication in the next generation. The next generation will be AI-native — AI embedded in the hardware and the software."
AI will deliver vastly improved performance, Vasishta said. "This is going to be required for the next generation to connect to the hundreds of billions of smart devices that will be prevalent in the next generation of wireless networks."
AI RAN partners
To boost research and development of AI-native wireless network hardware, software and architecture for 6G, Nvidia unveiled a partnership Tuesday with T-Mobile, Mitre, Cisco, ODC (a portfolio company of Cerberus Capital Management) and Booz Allen Hamilton. They're developing an AI-native wireless stack based on the Nvidia AI Aerial platform, which provides software-defined radio access networks (RANs) on the Nvidia accelerated computing.
The goal is to develop AI-RAN technology further, using AI to improve radio networks and deploy AI applications at the edge — running network and AI workloads together.
"The end goal is to have a fully integrated AI-native wireless network that sets new benchmarks in spectral efficiency, power efficiency, operational efficiency, security, cost-effectiveness and new opportunities for revenue generation," Vasishta said. "This will be a scalable solution ready for global deployments."
Sounds great — but it's unclear whether AI-RAN will prove practical in the real world. Intel VP and general manager Alex Quach, for one, is skeptical; he told Fierce a few weeks ago that GPUs require just too much power to implement at RAN sites.
Industry analyst Dean Bubley, founder and director of Disruptive Analysis, is also skeptical. He laid out arguments against AI-RAN in a lengthy LinkedIn post that's worth reading — it boils down to the argument that AI-RAN will prove overwhelmingly complex to implement. In addition to prohibitive power requirements, software engineers will oppose AI-RAN for security reasons, and RAN engineers will oppose AI-RAN as possibly disruptive to the network.
Still, everything is impossible in engineering until it isn't – so AI-RAN is worth watching.
Up in the air
Nvidia launched new additions to the Aerial Research portfolio Tuesday, a suite of Nvidia research tools for developing, training, simulating and deploying wireless networks. Upgrades include the Omniverse Digital Twin Service, for simulating 5G and 6G infrastructure from single towers to entire cities; the Aerial Commercial Test Bed on Nvidia MGX; and upgrades to Sionna 1.0 5G and 6G physical layer research software.
Nvidia has achieved 40% performance improvement in network simulations using the new tools, and expects those improvements to translate to the real world, Vasishta said.
Agentic AI for telco networks
Nvidia also announced it is working with partners on developing large telco models (LTMs) and AI agents custom-built for the telco industry to automate telco operations.
Of course, you're familiar with large language models (LLMs): for example, Google Gemini and the GPT models from OpenAI implemented in ChatGPT. LLMs are general-purpose models trained on the entire internet. Large telco models are trained specifically on telco native data, and together with AI agents, they automate network configurations, improving operational efficiency, boosting employee productivity and enhancing network performance.
SoftBank and Tech Mahindra have built new LTMs and AI agents, and Amdocs, BubbleRAN and ServiceNow are implementing the tools.
These tools can reconfigure networks in minutes, where these operations formerly took days, Vasishta said. He showed a demo of two agentic AIs communicating in a chat window to reconfigure a wireless network to serve a baseball stadium during a big game; the AIs working together were able to achieve a 30% simulated performance and throughput improvement on game day.
But AI factories were the main focus
Other than the cute robots, AI factories were the stars of the GTC keynote. As explained in a recent free report from Fierce Network Research (co-sponsored by Supermicro and Nvidia), AI factories are a new generation of data centers. Where conventional data centers run a variety of workloads, an AI factory is optimized for a single workload: running AI. Where a conventional factory produces cars or shoes, an AI factory produces AI tokens.
"They are called AI factories because they have one job, and one job only — generating these components [tokens] that we then reconstitute into music, into words, into videos, into research, into chemicals, proteins — reconstitute it into all kinds of information," Huang said.
While generative AI has resulted in an explosion of demand for compute, agentic AI and reasoning AI will require even more than that — massively more, Huang said at the keynote Tuesday. The reason is simple: With generative AI, you ask a question and get an answer — "blurting it out," Huang said — but with agentic AI and reasoning AI, the process is iterative, as the AI refines its results step-by-step, and each step requires more tokens.
"The amount of computation we have to do for inference is dramatically higher than we needed previously," Huang said.
To be sure, Nvidia estimates that the capital expenditure on the world's data centers, including cloud service providers and enterprises, will be $1 trillion through 2028, Huang said.
As noted in our report, Supermicro and Nvidia see AI factories as an opportunity for telcos, who are well positioned to implement these specialized data centers to offer infrastructure-as-a-service that meets global regulatory requirements for AI sovereignty. More than a dozen telcos worldwide are already implementing AI factories based on Nvidia blueprints.
At GTC Tuesday, Nvidia introduced several technologies to advance AI factory development.
Nvidia Dynamo is open source inference software for scaling AI reasoning models in AI factories at the lowest cost and the highest efficiency, orchestrating inference communications across thousands of GPUs. Dynamo doubles the performance and revenue of AI factories serving Lama models on today's Nvidia Hopper platforms, and it can boost the performance of tokens generated using DeepSeek-R1 by over 30x per GPU, Nvidia says.
Nvidia Omniverse Blueprint is a digital twin platform for designing and operating AI factories, testing and optimizing power, cooling and networking long before construction starts.
And Spectrum-X Photonics is a new line of networking switches to scale AI factories to millions of GPUs, providing 1.6 terabits per second per port to deliver 3.5x energy savings and 10x resilience.
About those cute robots
"Never work with children or animals" is an admonition attributed to the comedian W.C. Fields. If robots had been around a century ago, Fields would likely have added robots to the rule because those machines nearly stole the show at the GTC keynote.
In a video presentation, Nvidia displayed simulations of humanoid robots working in warehouses and factories, kitchens washing dishes side-by-side with a human, and — in a grand finale — a little robot that resembled the robot from the movie Wall-E scratched along in what appeared to be a desert, or the sands of Mars.
Then the Wall-E-looking robot took the stage with Huang, in real life, and the crowd leapt to its feet with their phones in the, squealing with delight (which is why we didn't get a photo — we couldn't get our phones over everyone's heads. Also, we were busy going, "Squee!").
The discussion of robotics wasn't just about cuteness. Far from it. Nvidia sees robots, self-driving cars and other autonomous machines — which it calls "physical AI" — as a significant growth area for the company and a multi-trillion-dollar global industry, meeting a critical global labor shortfall. "This could very likely be the largest industry of all," Huang said.
To drive humanoid robotic development, Nvidia launched several technologies, including Nvidia Isaac GROOT N1, which Nvidia touts as an open, customizable foundation model for generalized humanoid reasoning and skills, along with simulation frameworks and blueprints, such as the Nvidia Isaac GROOT Blueprint for generating synthetic data, as well as Newton, an open source physics engine under development with Google DeepMind and Disney Research, purpose-built for developing robots.
And that's a wrap for day one of the GTC conference. We'll be on the ground Wednesday and Thursday, bringing you the latest news and insights. And maybe we'll be able to get a selfie with that cute little robot.