NVIDIA's Blackwell Architecture: 25x More Performance for Next-Gen AI and Robotics

Introduction

At the annual GPU Technology Conference (GTC) 2025, often referred to as the "Super Bowl of AI," NVIDIA CEO Jensen Huang delivered a comprehensive keynote that outlined the company's vision for the next generation of AI and accelerated computing. Speaking to a packed stadium without a teleprompter or script, Huang showcased NVIDIA's latest innovations and strategic roadmap in what has become the premier event for the AI industry.

This year's GTC comes at a pivotal moment in computing history, as we witness the transition from generative AI to agentic AI and physical AI. Huang emphasized that the computational requirements for these advanced AI systems are approximately 100 times greater than anticipated just a year ago, creating both challenges and opportunities for the entire computing ecosystem.

Key Points

• The Blackwell architecture delivers 25x more performance than Hopper within the same power envelope, revolutionizing AI inference capabilities
• NVIDIA introduced Dynamo, a new operating system for AI factories that optimizes workload distribution across GPUs
• Major partnerships announced with GM for autonomous vehicles, T-Mobile and Cisco for edge computing
• New DGX Spark and DGX Station platforms bring AI development capabilities to enterprises and researchers
• NVIDIA presented a clear multi-year roadmap with Blackwell Ultra, Vera Rubin, and Rubin Ultra architectures
• For robotics, NVIDIA unveiled Isaac Groot N1 and Newton physics engine for physical AI development

The Evolution of AI and Computing

Huang began by reflecting on AI's remarkable journey over the past decade. Starting with perception AI for computer vision and speech recognition, the field evolved into generative AI that can translate between different modalities. Now, we're entering the era of agentic AI, where systems can perceive, reason, plan, and take action independently.

"Agentic AI basically means that you have an AI that has agency," Huang explained. "It can perceive and understand the context of the circumstance, it can reason very importantly about how to answer or how to solve a problem, and it can plan and action."

Beyond agentic AI, Huang highlighted the emergence of physical AI that understands concepts like friction, inertia, and object permanence—the foundation for the next wave of robotics.

The Three Scaling Laws of AI

Huang identified three fundamental challenges that have shaped AI development:

1. Solving the data problem - Finding ways to provide AI with sufficient digital experience to learn from
2. Training without human limitations - Enabling AI to learn at superhuman rates without being constrained by human supervision
3. Establishing effective scaling laws - Creating algorithms where more resources consistently yield smarter AI

"The computation requirement, the scaling law of AI, is more resilient and in fact hyper-accelerated," Huang noted. "The amount of computation we need at this point as a result of agentic AI, as a result of reasoning, is easily a hundred times more than we thought we needed this time last year."

This increased computational demand stems from reasoning capabilities that require AI to generate many more tokens as it works through problems step by step, checking its work and considering multiple approaches—similar to human thinking processes.

AI Factories: The New Computing Paradigm

Huang introduced the concept of "AI factories"—data centers purpose-built for generating tokens, the building blocks of AI. This represents a fundamental shift from retrieval-based computing to generative computing.

"The world is going through a transition in not just the amount of data centers that will be built, but also how it's built," Huang said. "Everything in the data center will be accelerated."

To illustrate this transformation, Huang shared that the capital expenditure for data centers is expected to reach $1 trillion by 2030, with the majority of this growth in accelerated computing infrastructure.

NVIDIA Blackwell: The Next Generation Architecture

The star of the keynote was undoubtedly the Blackwell architecture, which Huang described as a "giant leap" in inference performance. Blackwell represents a fundamental transition in computer architecture, moving from integrated MVLink to disaggregated MVLink, from air cooling to liquid cooling, and dramatically increasing component density.

"Our goal is to do scale up," Huang emphasized, showing how the new Grace Blackwell MVLink 72 rack essentially functions as a single massive GPU with 30 trillion transistors, 570 terabytes per second of memory bandwidth, and exaflop-level computing capability.

What makes Blackwell particularly remarkable is its performance for inference workloads. Huang demonstrated that Blackwell delivers 25x more performance than Hopper within the same power envelope, and up to 40x more performance for reasoning models.

"When Blackwell starts shipping in volume, you couldn't give Hoppers away," Huang quipped, highlighting the dramatic performance improvement.

NVIDIA Dynamo: The Operating System for AI Factories

To manage the complexity of these advanced AI systems, NVIDIA introduced Dynamo, an open-source operating system for AI factories. Dynamo optimizes workload distribution across GPUs, managing pipeline parallelism, tensor parallelism, expert parallelism, and disaggregated inferencing.

"Dynamo is essentially the operating system of an AI factory," Huang explained. "Whereas in the past, the way that we ran data centers, our operating system would be something like VMware... in the future, the application is not enterprise IT, it's agents, and the operating system is not something like VMware, it's something like Dynamo."

The name Dynamo references the machine that sparked the industrial revolution of energy—a fitting analogy for NVIDIA's vision of transforming computing.

Enterprise AI and New Development Platforms

Recognizing that AI needs to extend beyond cloud data centers, Huang introduced new platforms designed for enterprise computing:

1. DGX Spark - A compact AI development system with 20 CPU cores, 128GB of GPU memory, and one petaflop of computation power, priced at $30,000
2. DGX Station - A liquid-cooled personal workstation featuring Grace Blackwell architecture with 20 petaflops of performance and 72 CPU cores

"This is what a PC should look like," Huang said of the DGX Station. "This is the computer of the age of AI. This is what computers should look like, and this is what computers will run in the future."

These systems will be available through OEM partners including HP, Dell, Lenovo, and ASUS.

Strategic Partnerships Across Industries

Huang announced several major partnerships that will extend NVIDIA's AI technology across different sectors:

• General Motors selected NVIDIA to build their future self-driving car fleet, with AI capabilities for manufacturing, enterprise operations, and in-vehicle systems
• Cisco, NVIDIA, T-Mobile, CUS, and ODC will build a full stack for radio networks in the United States, bringing AI to telecommunications infrastructure
• Enterprises including Accenture, AMD, AT&T, BlackRock, Cadence, Capital One, DELL, EY, NASDAQ, SAP, and ServiceNow are integrating NVIDIA technology into their AI frameworks

The Future Roadmap: Beyond Blackwell

In a rare move for the technology industry, Huang laid out NVIDIA's multi-year roadmap to help customers plan their AI infrastructure investments:

• Blackwell Ultra (second half of 2025) - MVLink 72 with 1.5x more FLOPs, 1.5x more memory, and 2x more networking bandwidth
• Vera Rubin (second half of 2026) - MVLink 144 with a new CPU that's twice the performance of Grace, new GPU (CX9), new networking, and HBM4 memory
• Rubin Ultra (second half of 2027) - MVLink 576 with extreme scale-up capabilities, 15 exaflops of performance, and 4,600 terabytes per second of scale-up bandwidth

"We planned this out in multiple years," Huang explained, "because this isn't like buying a laptop. This isn't discretionary spend. This is spend that we have to go plan on."

Networking and Photonics: Enabling Massive Scale

To connect these increasingly powerful systems, NVIDIA announced breakthroughs in networking technology:

• Spectrum-X - NVIDIA's Ethernet technology that brings InfiniBand-like qualities to Ethernet networks, now being integrated into Cisco products for enterprise AI
• NVIDIA Photonics - The world's first 1.6 terabit per second co-packaged optical system using micro-ring resonator modulator (MRM) technology

The silicon photonics innovation is particularly significant, as it addresses the energy and cost challenges of scaling to millions of GPUs. Traditional transceivers would consume 180 watts per GPU and cost $6,000 per GPU at scale, while NVIDIA's photonic solution dramatically reduces both power consumption and cost.

Physical AI and Robotics

In the final segment of his keynote, Huang turned to robotics, highlighting how AI is enabling a new generation of autonomous systems. With a projected global shortage of 50 million workers by the end of the decade, robotics represents an enormous market opportunity.

NVIDIA introduced two significant technologies for robotics:

1. Isaac Groot N1 - An open-source foundation model for humanoid robots with a dual system architecture for "thinking fast and slow"
2. Newton - A physics engine developed in partnership with DeepMind and Disney Research, designed specifically for training robots with fine-grain rigid and soft body physics

"Physical AI and robotics are moving so fast," Huang urged. "Everybody pay attention to this space. This could very well likely be the largest industry of all."

Conclusion

Jensen Huang's GTC 2025 keynote painted a comprehensive picture of NVIDIA's vision for the future of computing. From the revolutionary Blackwell architecture to the long-term roadmap extending to Rubin Ultra, from enterprise AI platforms to advanced robotics capabilities, NVIDIA is positioning itself at the center of the AI transformation across every industry.

The keynote underscored several key themes: the exponential growth in computational requirements for advanced AI, the transition from data centers to AI factories, the importance of scale-up before scale-out in system architecture, and the emergence of physical AI as the next frontier.

As Huang concluded, "We have three AI infrastructures we're building: AI infrastructure for the cloud, AI infrastructure for enterprise, and AI infrastructure for robots." With this multi-pronged strategy and clear technology roadmap, NVIDIA is setting the pace for the next era of computing.

The question now is not whether AI will transform industries—it's how quickly organizations can adapt to this new computing paradigm and leverage these powerful new capabilities to solve the world's most challenging problems.

For the full conversation, watch the video here.