- Nvidia unveiled the Vera Rubin platform — seven new chips in production — promising 10x inference throughput per watt and one-tenth the cost per token versus Blackwell
- Nvidia and T-Mobile are pushing physical AI to the network edge via AI-RAN, turning base stations into AI inference infrastructure
- The open-source OpenClaw agent framework and Dynamo 1.0 inference OS are giving enterprises and telcos a production-ready stack for agentic AI deployment
NVIDIA GTC, SAN JOSE, CALIF. — Nvidia unveiled sweeping AI infrastructure upgrades this week, signaling a fundamental shift in how its AI compute is built, priced and deployed, which will impact telecommunications companies and the enterprises they serve.
The launches span new silicon, software and partnerships: a next-generation AI supercomputer platform, an open-source inference operating system, an enterprise agentic AI framework and a network edge AI push with T-Mobile. Taken together, they amount to Nvidia's most comprehensive effort yet to own every layer of the AI stack — from data center silicon to the radio access network.
"Inference is the engine of intelligence, powering every query, every agent and every application," Jensen Huang, Nvidia founder and CEO, said in a two-and-a-half hour keynote Monday. "With Nvidia Dynamo, we've created the first-ever operating system for AI factories. The rapid adoption across our ecosystem shows this next wave of agentic AI is here, and Nvidia is powering it at global scale."
Here are the top takeaways for the telecom and enterprise networking community.
Vera Rubin raises the bar on AI factory economics
The headline announcement was the Vera Rubin platform — a full-stack AI supercomputer comprising seven new chips across five rack-scale systems, all in production now. The platform includes the Rubin GPU, Vera CPU, Groq 3 LPU inference accelerator, BlueField-4 STX storage system and Spectrum-6 Ethernet switch, all co-designed to operate as a single coherent system.
The platform is built to vastly improve economics for AI inference. The Vera Rubin NVL72 rack — integrating 72 Rubin GPUs and 36 Vera CPUs connected via NVLink 6 — delivers up to 10x higher inference throughput per watt at one-tenth the cost per token compared with the previous Blackwell generation. Huang framed this as a revenue inflection for any organization running AI at scale: The same power envelope can now generate dramatically more tokens, translating directly to either lower costs or higher service tiers, enabling organizations to charge more for better quality of service.
The Groq 3 LPU integration is notable. In December, Nvidia inked a $20 billion licensing deal with Groq, whose language processing units were built from the ground up for inference workloads. That bet is now paying off: By pairing Groq LPUs with Rubin GPUs, Nvidia delivers a 35x performance boost at the highest inference tier, Huang said. Cloud providers and telcos building AI factory infrastructure can extract significantly more value from fixed power capacity — a critical constraint at a time when data center power availability is a primary bottleneck. Vera Rubin systems will be available from AWS, Google Cloud, Microsoft Azure, Oracle Cloud and Nvidia cloud partners including CoreWeave and Vultr starting in the second half of 2026.
Dynamo 1.0 brings an inference operating system to AI factories
Dynamo 1.0 is open-source software that functions as a distributed operating system for AI inference workloads, orchestrating GPU and memory resources across a cluster and routing requests intelligently based on workload type, context length and performance objectives. In industry benchmarks, it boosted inference performance on Blackwell GPUs by up to 7x, Nvidia said.
As agentic AI systems move to production, inference has become a complex resource orchestration challenge. Requests of varying sizes and modalities arrive in unpredictable bursts, and managing them efficiently determines both cost and quality of service. Dynamo is free and open source, integrating natively with popular frameworks including LangChain, vLLM and SGLang. Adoption is already broad: AWS, Azure, Google Cloud and Oracle have integrated it, as have enterprises including BlackRock, ByteDance, PayPal and Pinterest.
T-Mobile partnership brings physical AI to the network edge
Nvidia and T-Mobile bowed a partnership to integrate physical AI applications on AI-RAN-ready infrastructure — effectively turning base stations into AI inference platforms at the network edge. For the full story, see our Monday story from GTC.
Network automation has become the top AI use case for telco ROI, and AI-RAN represents the next step in that evolution, embedding GPU-accelerated compute directly into the radio access network.
OpenClaw and the agentic enterprise
Nvidia announced NemoClaw, an enterprise-ready reference stack built on OpenClaw, the open-source agentic AI framework released by developer Peter Steinberger in November that Huang called the fastest-growing open-source project in history. NemoClaw layers enterprise security via a new OpenShell runtime onto OpenClaw, adding policy enforcement, network guardrails and a privacy router to make autonomous AI agents safe for corporate deployment.
Huang compared OpenClaw's significance to Linux, HTML and Kubernetes — a foundational layer that every company will need a strategy for. Every SaaS company will become an "agentic AI as a service" company, and every enterprise will eventually issue token budgets to employees the way it issues software licenses today, Huang said. For telcos, which operate some of the most compliance-sensitive IT environments in the world, the addition of enterprise-grade security controls to an open-source agent framework is a meaningful development.
The Nemotron Coalition expands open model access
The "Nemotron Coalition" sounds like a gang of villains from a superhero movie, but no — it's an alliance of AI labs, launched by Nvidia, including Mistral AI, Perplexity, LangChain, Black Forest Labs, Cursor and Thinking Machines Lab, to co-develop open frontier models. The first output will be a base model co-developed by Mistral AI and Nvidia, trained on DGX Cloud and open-sourced as the foundation for the upcoming Nemotron 4 model family.
Open frontier models give enterprises and telcos an alternative to relying solely on proprietary models from OpenAI, Anthropic or Google. For telcos pursuing sovereign AI strategies — particularly outside the US — the ability to fine-tune and host a frontier-quality base model represents a meaningful new option.
Physical AI moves off the drawing board
Nvidia announced broad robotics ecosystem expansions, including new Cosmos world models, Isaac simulation frameworks and GR00T robotics foundation models. GR00T N1.7 is now available in early access with commercial licensing. New autonomous vehicle partners — BYD, Hyundai, Kia, Nissan and Geely — joined the Nvidia DRIVE ecosystem, adding roughly 18 million vehicles per year to the platform.
For telco customers in manufacturing, logistics and transportation, the maturation of Nvidia's physical AI stack means production-ready robotic systems are arriving faster, with lower integration costs. For telcos themselves, industrial AI deployments are a growing driver of both edge compute demand and private wireless network adoption.