GTC 2026 Telecom: The AI Grid

Category: Telecom Special Address · Year: 2026 · ▶ Watch

Switch language → 中文

Segments (21)

  • 00:00 · Introduction to the AI Grid
    • Ronnie Vasishta introduces the concept of the AI Grid, a new infrastructure buildout connecting the physical world with AI.
  • 03:24 · Three Tectonic Shifts in Telecom
    • Discussion of accelerated computing, the evolution of AI applications, and the transition from 5G to 6G.
  • 05:57 · Merging the Telecom and AI Stacks
    • How accelerated computing enables the convergence of purpose-built telecom stacks with general-purpose AI stacks.
  • 07:55 · Telcos’ Prime Locations for AI
    • Highlighting the unique advantage of telecom operators in possessing distributed land, power, and shell infrastructure.
  • 10:15 · Sovereign AI Factories
    • Overview of 24 telcos globally building sovereign AI factories to serve local regions and enterprises.
  • 11:25 · Distributed Edge Primed for AI
    • The massive untapped opportunity of utilizing spare power at regional POPs and central offices for AI compute.
  • 12:20 · New AI Applications and Tokenomics
    • Exploring the killer apps for the edge and the new economic model based on token generation and latency.
  • 15:55 · Personal AI Edge Inference Demo
    • A demonstration comparing the latency and cost of running an AI assistant in the cloud versus at the edge.
  • 17:30 · Decart Edge Inference Demo
    • A demonstration showing the performance benefits of edge rendering for real-time video personalization.
  • 19:10 · AT&T and Cisco IoT AI Grid
    • Announcement of AT&T and Cisco building an AI Grid for IoT using NVIDIA RTX 6000 GPUs.
  • 20:45 · Comcast AI at the Edge
    • Comcast’s initiative to bring AI to the edge for personalized advertising and small business agents.
  • 21:40 · Spectrum AI Grid for Movie Makers
    • Spectrum’s deployment of low-latency hubs for high-performance graphics rendering in the media industry.
  • 22:35 · Akamai Intelligent AI Orchestration
    • Akamai’s global-scale implementation of AI inference orchestration across thousands of edge locations.
  • 23:25 · RAN as a Workload on the AI Grid
    • The shift towards AI-RAN, allowing radio access networks to run as software workloads alongside AI applications.
  • 25:30 · T-Mobile Physical AI Apps on AI-RAN
    • T-Mobile’s integration of physical AI apps, like traffic monitoring, onto their AI-RAN infrastructure in San Jose.
  • 28:25 · Indosat Ooredoo Hutchison Sovereign AI Factory
    • IOH’s efforts to build an AI Grid for local innovation and close the digital divide in Indonesia.
  • 30:05 · NVIDIA AI Grid Reference Architecture
    • Introduction of the full-stack reference architecture and ecosystem partners for building the AI Grid.
  • 31:40 · Aerial Omniverse Digital Twin Demo
    • A video presentation showcasing the simulation of AI-native wireless networks in a digital twin environment.
  • 33:25 · 6G as the Fabric for Physical AI
    • The role of 6G in connecting billions of devices, robots, and AI agents in the physical world.
  • 36:45 · Democratizing 6G Research
    • NVIDIA’s initiatives and developer programs aimed at accelerating 6G research and development.
  • 39:55 · Road to the AI-Native Telco
    • Concluding remarks on the generational opportunity for telecom operators to embrace AI-native architectures.

Product Announcements (3)

  • [30:05] NVIDIA AI Grid Reference Architecture
    • A full-stack hardware and software reference architecture for deploying AI and RAN workloads on shared infrastructure.
    • specs: Integrates AI Grid Control Plane, NVIDIA AI-RAN Computer, and supports various ecosystem partners.
    • availability: Available through ecosystem partners.
  • [30:05] RTX Pro 6000 Blackwell Server Edition
    • A server-class GPU designed for AI-RAN and edge AI workloads.
    • specs: Optimized for distributed edge deployments and AI inference.
    • availability: Not specified.
  • [31:40] Aerial Omniverse Digital Twin
    • A platform for simulating and optimizing wireless networks and physical AI agents.
    • specs: Physically accurate 3D environments, ray-tracing channel simulation, and integration with 3rd party digital twins.
    • availability: Not specified.

Specific Numbers (12)

Timestamp Metric Value Context
10:21 Telco AI Factories 24 Number of telcos building Sovereign AI Factories across 5 continents.
11:25 Regional POPs 100,000 Estimated number of regional points of presence globally.
11:25 Spare Energy 1 MW Average spare energy available per regional POP site.
11:25 AI Capacity 100 GW Total estimated AI capacity available today at distributed edge locations.
16:45 Latency 400ms End-to-end latency for the Personal AI app running on an Edge SLM, compared to 2000ms on a Cloud LLM.
16:55 Cost per 1M Tokens $0.02 Cost of running the Personal AI app on an Edge SLM, compared to $0.80 on a Cloud LLM.
18:40 Frames Per Second (FPS) 30 FPS achieved by Decart’s video personalization app at the edge, compared to 7-16 FPS in the cloud.
20:57 Edge Locations 65 million Number of locations reached by Comcast’s edge compute network.
22:48 Edge and Core Locations 4,400+ Number of locations in Akamai’s global AI inference network.
27:23 Performance Improvement 5x Improvement in speed of seeing actions at a traffic intersection using T-Mobile’s AI-RAN infrastructure.
37:20 Developer Program Members 7,000+ Number of members in the NVIDIA 6G Developer Program.
37:38 Sionna Downloads 375,000+ Number of downloads for the Sionna link-level simulator.

Benchmark Claims (4)

  • [16:45] End-to-End Latency (Personal AI): 400ms
    • vs: 2000ms (Cloud LLM)
    • gain: 5x reduction in latency for a more responsive conversational experience.
  • [16:55] Cost per 1M Tokens (Personal AI): $0.02
    • vs: $0.80 (Cloud LLM)
    • gain: 40x reduction in inference cost by utilizing edge SLMs.
  • [18:40] Frames Per Second (Decart Video Personalization): 30 FPS
    • vs: 7-16 FPS (Cloud)
    • gain: Smooth, real-time video rendering by mitigating cloud jitter through edge proximity.
  • [27:23] Vision AI Processing Speed (T-Mobile/San Jose): 5x faster
    • vs: Previous non-edge implementations
    • gain: Significantly faster anomaly detection at traffic intersections.

Customer Stories (6)

  • [19:15] AT&T and Cisco
    • Teamed up with NVIDIA to build an AI Grid for IoT, deploying RTX 6000 GPUs in UCS servers.
    • outcome: Created a zero-trust, on-demand AI Grid to deliver intelligence at the cellular network edge for public safety and other use cases.
  • [20:45] Comcast
    • Partnered with Personal AI and Decart to bring AI to their massive edge compute network.
    • outcome: Enabled hyper-personalized ads and small business concierge agents with low latency.
  • [21:40] Spectrum
    • Transformed data centers into low-latency hubs using RTX 6000 GPUs.
    • outcome: Provided high-performance graphics rendering capabilities for movie makers requiring complex 3D collaboration.
  • [22:35] Akamai
    • Implemented a global-scale Intelligent AI Orchestration Platform across over 4,400 locations.
    • outcome: Enabled SLA-aware distribution of content and AI applications globally.
  • [25:30] T-Mobile
    • Integrated physical AI apps, like Metropolis VSS, onto their AI-RAN ready infrastructure in San Jose.
    • outcome: Achieved a 5x performance improvement in processing traffic intersection camera feeds for anomaly detection.
  • [28:25] Indosat Ooredoo Hutchison (IOH)
    • Built a Sovereign AI Factory and deployed AI applications like the Sahabat AI app to the distributed edge.
    • outcome: Provided localized AI services for education, agriculture, and healthcare to help close the digital divide in Indonesia.

Key Technologies (4)

  • AI Grid: A distributed infrastructure that connects the physical world and deploys AI at scale into the hands of end-users.
  • AI-RAN: Allows Radio Access Network (RAN) and AI workloads to run concurrently on the same software-defined, accelerated computing infrastructure.
  • Tokenomics: An economic model for AI inference that values compute based on metrics like tokens per second and time to first token.
  • Aerial Omniverse Digital Twin: A simulation platform that uses physically accurate 3D environments and ray-tracing to optimize wireless networks and physical AI agents.

Demos Shown (3)

  • [15:55] A side-by-side comparison of a Personal AI auto repair voice assistant running on a Cloud LLM versus an Edge SLM, highlighting latency and cost differences.
    • True
  • [17:30] A demonstration by Decart showing real-time video personalization (inserting a product into a video stream), comparing the frame rate of cloud rendering versus edge rendering.
    • True
  • [31:40] A video presentation of the Aerial Omniverse Digital Twin, showing how developers can simulate 6G networks, beamforming, and autonomous robots in a virtual city.
    • True

Predictions / Commitments (3)

  • [04:55, Next generation of telecom networks (6G)] 6G is being born in the era of AI and will be deployed to be the fabric of AI.
  • [06:55, Near future] The convergence of the compute stack will create completely new KPIs for the industry and lead to unicorns built overnight.
  • [33:40, 2-3+ years from now] 6G will provide the connectivity fabric required to connect hundreds of billions of humans, machines, and AI agents.

Companies Mentioned (7)

Cisco · Nokia · Ericsson · SoftBank · AWS · Fujitsu · Samsung

Notable Quotes (3)

AI is redefining computing and driving the largest infrastructure buildout in human history—and telecommunications is next. — Ronnie Vasishta (quoting Jensen Huang) @ 05:30

The ability to provide tokenomics, tokenization in that distributed network of value is very, very unique to this infrastructure. — Ronnie Vasishta @ 12:28

There has never been a better time to be in Telcos. — Ronnie Vasishta @ 40:49

Key Topics

AI Grid · Telecommunications · AI-RAN · Edge Computing · 6G Networks · Accelerated Computing · Tokenomics · Physical AI · Sovereign AI · Digital Twins · Network Orchestration · Software-Defined Infrastructure

Takeaways

  • The telecommunications industry is undergoing a massive transformation as its purpose-built infrastructure merges with the general-purpose AI compute stack.
  • Telcos possess a unique advantage with their highly distributed real estate, power, and connectivity, making them prime candidates to build the ‘AI Grid’.
  • Deploying AI at the edge (AI Grid) significantly reduces latency and costs compared to cloud-based LLMs, enabling new ‘killer apps’ like real-time video personalization and physical AI.
  • AI-RAN allows operators to run radio access networks as software workloads alongside AI applications on the same accelerated hardware, maximizing utilization.
  • Major telecom and tech companies (AT&T, T-Mobile, Comcast, Cisco, Nokia, Ericsson) are already actively deploying and testing AI Grid and AI-RAN architectures.
  • 6G will be the first network generation ‘born in the era of AI’ and will serve as the essential connectivity fabric for billions of physical AI agents and robots.
  • NVIDIA is providing full-stack reference architectures, GPUs (RTX Pro 6000), and simulation tools (Aerial Omniverse Digital Twin) to accelerate the telecom industry’s transition to AI-native networks.