Chips, Robots, Agents: 5 Big AI Bets Jensen Huang Announced at NVIDIA GTC 2026
By Sanjana Gupta ·
NVIDIA CEO Jensen Huang used the company’s GTC keynote to outline a full-stack shift towards what he called the next phase of AI, centred on inference, agents, and physical systems. The focus has shifted from training models to running them continuously for inference, agents, and real-world systems.
The announcements span chips, software, telecom, robotics and creative tools, signalling a move beyond training models to running them at scale. Huang also said demand for AI infrastructure could reach $1 trillion by 2027, reflecting what he described as an inflexion point for inference.
Here is a breakdown of what was announced in this ‘Super Bowl of AI’.
The Hardware Push
NVIDIA’s hardware announcements reflect the shift that data centres are no longer storage systems. They are becoming factories that generate tokens, the basic unit of AI output. The company introduced the Vera Rubin platform not as a chip, but as a full system designed to run AI factories at scale.
“Vera Rubin is a generational leap—seven breakthrough chips, five racks, one giant supercomputer—built to power every phase of AI,” said Huang.
The platform combines CPUs, GPUs, networking, and storage into a tightly integrated architecture optimised for inference workloads. Huang positioned the platform as a step towards AI factories, where compute infrastructure produces intelligence at scale rather than just training models.
Huang framed it as a departure from earlier generations. Instead of selling individual processors, NVIDIA is now building “one giant computer” designed for agentic systems that must think, act, and respond in real-time.
NVIDIA launched the Vera CPU, purpose-built to orchestrate AI agents and handle memory-intensive workflows. Unlike traditional CPUs, it focuses on coordination and data movement, which Huang described as critical for systems that rely on both structured and unstructured data. This marks a deeper move into CPU territory, an area historically dominated by Intel and AMD.
The chip focuses on single-thread performance and orchestration, supporting systems that run multiple AI agents simultaneously.
The company also pushed its Dynamo software stack into production, positioning it as the operating system of AI factories. Dynamo manages inference workloads across clusters, optimising token generation and distribution.
This comes just as the rest of the chip players and the industry move towards full rack-scale solutions rather than just procuring chips, showcasing that hardware alone is no longer enough. It must be tightly coupled with software to deliver performance gains in inference, where latency, throughput, and energy efficiency define value.
Huang described this layer as essential to scaling AI beyond experimentation and into production environments.
Agents, Factories, Twins, and Infrastructure
Agentic AI emerged as the keynote’s dominant theme. NVIDIA introduced frameworks, models, and tools that enable AI systems to plan, reason, and execute tasks autonomously. The company positioned agents as the next computing interface, replacing traditional apps with autonomous workflows.
Huang described this as the next stage of AI evolution. Systems have moved from perception to generation, then to reasoning, and now to action. “An AI that could reason now has become an AI that can actually do work,” he said.
To support this shift, NVIDIA launched NemoClaw, a stack designed to secure and manage AI agents with privacy controls and guardrails. It builds on the broader OpenClaw ecosystem, which Huang described as foundational for agent-based computing.
OpenClaw is a long-running AI assistant capable of taking actions independently. It can be hosted on the user’s local hardware.
It also expanded its open model ecosystem through initiatives such as the Nemotron Coalition, which aims to build open frontier models and accelerate adoption across industries. With this, the company is betting that open ecosystems will accelerate adoption across industries.
At the infrastructure level, NVIDIA introduced the DSX AI Factory reference design and Omniverse-based digital twin blueprints. These allow companies to simulate AI factories before deploying them, reducing build time and improving efficiency.
The idea is to treat AI infrastructure like industrial systems, designed, tested, and scaled systematically.
NVIDIA also launched the BlueField-4 STX storage architecture to support AI workloads. It focuses on moving and managing data efficiently across systems. This is critical for inference, where speed and latency matter more than raw compute.
Telecom, Space, Industrial AI Enter the Mix
Beyond core infrastructure, NVIDIA is extending AI into existing global systems. The company announced partnerships with telecom players, including T-Mobile, to integrate AI into radio access networks (AI-RAN), turning base stations into distributed compute platforms.
Huang said this infrastructure, once built for connectivity, will now run AI workloads at the edge. The shift could allow real-time applications to operate directly within network environments.
NVIDIA also introduced its space computing initiatives, positioning satellites and orbital data centres as extensions of AI infrastructure. These systems aim to process data closer to its source, reducing latency and enabling global-scale AI operations.
In industrial sectors, the company is working with software providers to bring AI into design, engineering, and manufacturing workflows. These systems rely on simulation, digital twins, and real-time data processing, aligning with NVIDIA’s broader push towards AI-driven industrial systems.
Physical AI, Robotics, and Autonomy
A major part of NVIDIA’s vision centres on physical AI, referring to systems that operate in the real world, combining perception, reasoning, and action. The company introduced a physical AI data factory blueprint, designed to generate and process training data for robotics, autonomous vehicles, and vision systems.
NVIDIA also expanded its robotics ecosystem, working with partners like ABB Robotics and Figure AI to deploy AI systems in real-world environments rather than simulations. Other partners include AGIBOT, Agility, FANUC, Hexagon Robotics, KUKA, Skild AI, Universal Robots, World Labs, and YASKAWA. These include humanoid robots and industrial automation systems powered by vision-language-action models.
In autonomous driving, NVIDIA deepened partnerships with global automakers, including BYD, Geely, Isuzu, and Nissan, and expanded its DRIVE Hyperion platform for level 4 autonomy. These systems aim to handle complex driving scenarios without human intervention in certain geofenced areas.
NVIDIA’s full-stack robotaxis will also be launched with Uber across 28 markets by 2028, beginning with Los Angeles and San Francisco in the first half of 2027.
It also announced that it is working with global industrial software leaders such as Cadence, Siemens, NXP, Synopsys, and others to bring CUDA-X, Omniverse, and GPU-accelerated industrial software and tools to global players.
These include FANUC, HD Hyundai, Honda, JLR, KION, Mercedes-Benz, MediaTek, PepsiCo, Samsung, SK hynix, and TSMC to accelerate design, engineering, and manufacturing.
Creative Tools and Gaming
The chip giant also extended its AI strategy into creative and consumer domains. The company announced a partnership with Adobe to develop agentic workflows for design and marketing, allowing AI systems to automate parts of content creation and production.
In gaming, NVIDIA introduced DLSS 5, which uses neural rendering to combine traditional graphics with generative AI. This approach blends structured data from 3D environments with probabilistic AI models to create more realistic visuals.
Huang described this fusion as a broader pattern that will repeat across industries, where structured and unstructured data systems converge to power AI applications.
NVIDIA is shifting beyond just training AI models. Huang framed this as the “Age of Inference,” where real-time AI becomes the dominant workload.
This shift is also economic. AI factories produce tokens, which translate into services, applications, and revenue. Token efficiency will now define the next phase of competition.