NVIDIA's Blackwell Is Powering the AI Token Factory of the Future

Updated: March 19 2025 19:34
NVIDIA's GTC (GPU Technology Conference), now described as the "Super Bowl of AI," showcased just how far we've come from traditional computing paradigms to the new era of AI-driven innovation.

Twenty-five years after the introduction of GeForce, NVIDIA continues to push boundaries, but with a dramatic pivot toward artificial intelligence. What started as a graphics company has transformed into the backbone of the AI revolution, providing the computational power needed for everything from autonomous vehicles to humanoid robots.


As CEO Jensen Huang eloquently put it during his keynote address: "GeForce brought CUDA to the world. CUDA enabled AI, and AI has now come back to revolutionize computer graphics." This circular journey perfectly encapsulates how NVIDIA has not only adapted to technological shifts but actively driven them forward.




The progression of AI over the past decade has been nothing short of extraordinary. NVIDIA identifies three major waves in this evolution:

• Perception AI: The initial breakthrough came with computer vision and speech recognition systems that could interpret the world around them.
• Generative AI: The last five years have focused on teaching AI to translate between different modalities - text to image, image to text, text to video, amino acids to proteins, and properties to chemicals.
• Agentic AI: The newest advancement involves AI with agency - systems that can perceive, understand context, reason about problems, and plan actions. These AI agents can use tools, understand multimodal information, and apply newfound knowledge to complete tasks.

These advancements have fundamentally transformed computing from a retrieval model to a generative model. Rather than creating content in advance and fetching appropriate versions when needed, AI now understands context and meaning, generating customized responses in real-time.



One of the most critical insights from NVIDIA's presentation was the revelation about computation requirements. Jensen Huang explained that the world "got it wrong" in estimating how much computational power modern AI systems would need. The scaling law of AI is proving to be far more resilient and hyper-accelerated than anticipated.

The reasoning capabilities of modern AI require substantially more tokens and processing power - easily 100 times more than what experts thought necessary just a year ago. This is because reasoning AIs break down problems step by step, generating multiple tokens for each reasoning step rather than producing a single response in one shot.


This has created an enormous challenge for the computing industry, reflected in the dramatic increase in GPU shipments. NVIDIA showed that just one year after introducing their Blackwell architecture, shipments to the top four cloud service providers have already eclipsed the peak year of their previous Hopper architecture - and this doesn't even include AI companies, startups, or enterprise deployments.



At the heart of NVIDIA's current AI strategy is the Blackwell architecture, which represents a fundamental transition in computer architecture. Instead of simply scaling out with distributed computing (connecting many small computers), NVIDIA has focused on scaling up with more powerful individual systems before scaling out. The Blackwell system architecture features:

• Disaggregated MV-Link switches separated from the compute nodes
• Complete liquid cooling that enables compression of compute nodes
• 600,000 components per rack (compared to 60,000 per computer previously)
• One exaflop (a million trillion floating point operations per second) in a single rack

This architecture is particularly optimized for inference - the process of generating tokens in response to prompts. As Jensen explained, inference is token generation by a factory, and that factory directly affects quality of service, revenues, and profitability.


To manage these incredibly complex systems, NVIDIA announced Dynamo, which Jensen described as "the operating system of an AI factory." This software handles numerous complex tasks such as pipeline parallellism, tenso parallelism, workload management, etc.


Named after the machine that started the last industrial revolution, Dynamo is open source and already has numerous partners working with NVIDIA to implement it. Performance benchmarks showed that Blackwell with MV-Link 72 and Dynamo delivers 40 times the performance of Hopper for reasoning model inference.


In an unusual move for the tech industry, NVIDIA openly shared its product roadmap for several years ahead, allowing customers to plan their AI infrastructure investments with confidence:

• Blackwell Ultra (2nd half of 2024): 1.5x more FLOPS, 1.5x more memory, 2x more bandwidth
• Vera Rubin (2nd half of 2025): Named after the astronomer who discovered dark matter, featuring a new CPU with twice the performance of Grace, new GPU, CX9 networking, and MV-Link 144
• Rubin Ultra (2nd half of 2027): MV-Link 576 with extreme scale-up capabilities, 15 exaFLOPS (compared to Blackwell's 1 exaFLOP), and 4,600 terabytes per second scale-up bandwidth

To support this scaling, NVIDIA also announced breakthrough silicon photonics technology that will allow them to connect hundreds of thousands to millions of GPUs without the prohibitive power and cost requirements of traditional transceivers.





NVIDIA isn't limiting its focus to cloud data centers. The company is actively working to bring AI to enterprises, edge computing, and specialized industries.

For telecommunications, NVIDIA announced a partnership with Cisco, T-Mobile, Cerebrus, and ODC to build a full stack for radio networks in the United States, bringing AI to the edge. This is significant considering that $100 billion of the world's capital investments each year goes into radio networks and data centers provisioning for communication.


For enterprise computing, NVIDIA introduced the DGX Station and DGX Spark, powerful AI workstations that will be available through OEMs like HP, Dell, Lenovo, and ASUS. These workstations feature 20 petaFLOPS of computing power and are designed for data scientists and researchers worldwide.

NVIDIA is also revolutionizing storage systems, transitioning from retrieval-based storage to semantics-based storage that continuously embeds information in the background. This allows users to interact with their data through natural language rather than traditional retrieval methods.



Perhaps the most exciting frontier showcased at GTC was in robotics and physical AI. Jensen noted that by the end of this decade, the world will face a shortage of at least 50 million workers, creating an enormous opportunity for robotic systems. NVIDIA is addressing three core challenges in robotics:

• Solving the data problem
• Creating the right model architecture
• Establishing scaling laws for smarter AI

To tackle these challenges, NVIDIA has created Omniverse, their operating system for physical AIs, and added two critical technologies:

• Cosmos: A generative model that understands the physical world and can create infinite virtual environments for training robots
• Newton: A physics engine developed in partnership with google DeepMind and Disney Research, specifically designed for fine-grain rigid and soft body simulation, tactile feedback, and fine motor skills

The company also announced Groot N1, a generalist foundation model for humanoid robots featuring a dual system architecture inspired by human cognitive processing - a "slow thinking" system for perception and reasoning, and a "fast thinking" system for translating plans into precise robot actions.

In a significant move for the robotics community, NVIDIA announced that Groot N1 will be open-sourced, allowing developers worldwide to build upon this foundation.

Huang explained the need for “a physics engine that is designed for very fine-grain, rigid and soft bodies, designed for being able to train tactile feedback and fine motor skills and actuator controls.” The focus on simulation for robot training follows the same pattern that has proven successful in autonomous driving development, using synthetic data and reinforcement learning to train AI models without the limitations of physical data collection.



Jensen Huang described modern data centers as "AI factories" with one primary job: generating tokens that can be reconstituted into music, words, videos, research, chemicals, proteins, and countless other forms of information.

This transformation requires not just new hardware but entirely new ways of thinking about computing infrastructure. From the chip to the system architecture, from networking to storage, and from operating systems to programming models - everything is being reimagined for the AI era.


The future belongs to those who can build, manage, and optimize these AI factories effectively, and NVIDIA has positioned itself at the center of this technological revolution. With their comprehensive roadmap, open ecosystem approach, and relentless innovation, they are not just participating in the AI transformation - they're actively driving it forward.