NVIDIA CEO Jensen Huang announced a series of breakthrough advances in AI computing power at the company’s GTC’s March 2025 Keynote, describing what he called “a $1 trillion computing turning point.” The keynote reveals the production preparation of Blackwell GPU architecture, a multi-year roadmap for future architectures, major breakthroughs in AI networks, new enterprise AI solutions, and significant developments in robotics and physical AI.
“Token Economy” and AI Factory
At the heart of Huang’s vision is the concept of “tokens”, as the basic building block of AI, and the emergence of “AI factories” as a dedicated data center designed for generating computing.
Huang told the audience: “This is how intelligence is made, a new type of token generator, the foundation of AI. The token opens up a new frontier.” He stressed that tokens can “turn images into scientific data charts the atmosphere of aliens”, “decode the laws of physics” and “see disease before it lasts.”
This vision represents the transition from traditional “retrieval computing” to “generated computing” where AI understands the context and generates answers, rather than just getting pre-stored data. According to Huang, this transition requires a new type of data center architecture, namely, “the computer has become a generator of tokens, not file retrieval.”
Blackwell Architecture brings huge performance growth
NVIDIA BLACKWELL GPU building is now in “all production”, and the company claims to have “hopper performance of 40 times” in the inference model under the same power conditions. This architecture includes support for FP4 accuracy, resulting in significant energy efficiency improvements.
“ISO Power, Blackwell is 25 times,” Huang said, highlighting the huge efficiency improvements in the new platform.
The Blackwell architecture also supports extreme expansion through technologies such as NVLink 72, thereby enabling a large-scale unified GPU system. Huang predicts that Blackwell’s performance will enable the previous generation of GPUs to significantly reduce the ideal demanding AI workloads.
(Source: NVIDIA)
Predictable roadmap for AI infrastructure
NVIDIA outlines the regular annual pace of its AI infrastructure innovation, enabling customers to plan their investments more certainly:
- Blackwell Ultra (2025): Upgrade to Blackwell platform with increased FLOP, memory and bandwidth.
- Vera Rubin (second half of 2026): A new architecture with twice the performance of its CPU, a new GPU, as well as the next generation of NVLINK and memory technology.
- Rubin Ultra (2027): An extreme enlarged architecture designed to compute 15 exaflops per rack.
Democratic AI: From the Internet to the Model
To realize the vision of widespread adoption of AI, NVIDIA announced a comprehensive solution covering networks, hardware and software. At the infrastructure level, the company is addressing the challenge of connecting hundreds of thousands or even millions of GPUs through massive investments in silicon photonics technology. Their first co-package optical (CPO) silicon photonic system is a 1.6 Terabit CPO per second based on microring resonator modulator (MRM) technology, which promises to save a lot of power and density compared to traditional transceivers, making connections between large numbers of GPUs across different sites more efficient.
While building the foundation for large-scale AI factories, NVIDIA brings AI computing power to both individuals and smaller teams. The company has launched a new series of DGX personal AI supercomputers powered by the Grace Blackwell platform, designed to empower AI developers, researchers and data scientists. The lineup includes DGX Spark, a compact development platform and DGX station, a high-performance desktop workstation with liquid cooling, and an impressive 20 Petaflops Compute.
NVIDIA DGX SPARK (Source: NVIDIA)
NVIDIA announced the family of open llama models with reasoning capabilities in complementing the advancements in these hardware, which aims to prepare businesses for building advanced AI agents. These models are integrated into NVIDIA NIM (NVIDIA Inference Microservice), allowing developers to deploy them on various platforms from on-premises workstations to the cloud. This approach represents a full-stack solution adopted by enterprise AI.
Huang stressed that these initiatives have been enhanced by extensive collaboration with major companies in multiple industries that integrate NVIDIA models, NIMs and libraries into their AI strategies. This ecosystem approach is designed to accelerate adoption while providing flexibility for different enterprise needs and use cases.
Physical AI and Robotics: $50 trillion opportunity
According to Huang, NVIDIA views physical AI and robotics as “a $50 trillion opportunity.” The company announced the open source NVIDIA ISAAC GR00T N1, which is described as a “generalized basic model for humanoid robots.”
A significant update to the NVIDIA COSMOS World Foundation model provides unprecedented controls for robotic training using NVIDIA Omniverse. As Huang explains: “Using Omniverse to regulate the universe, the universe generates an infinite number of environments that allow us to create data that is rooted by us, control our but systematically infinite at the same time.”
The company has also partnered with Google Deepmind and Disney Research to develop a new open source physics engine called “Newton”. The engine is designed for high-fidelity robot simulations, including rigid and software, haptic feedback and GPU acceleration.
Isaac GR00T N1 (Source: NVIDIA)
Agent AI and industry transformation
Huang defines “agent AI” as AI, whose “agents” can “perceive and understand context,” “cause,” “plan and act,” and even use tools and learn from multimodal information.
“Agent AI basically means that you have an AI with an agency. It can be contextualized by perception and understanding the situation. It can reason, and it is very important to reason about how to answer or how to solve a problem, and it can be planned and acted. It can be planned and acted. It can use tools,” Huang explained.
This capability is driving a surge in computing demand: “The amount of calculation required, the scaling law of AI is more resilient, and is actually super-accelerated. Due to the results of reasoning, the amount of calculations we need is a hundred times more than we needed during this period last year,” he added.
Bottom line
Jensen Huang’s GTC 2025 Keynote presents a comprehensive vision for the AI-driven future, characterized by smart agents, autonomous robots and dedicated AI factories. NVIDIA’s announcement on hardware architecture, networking, software and open source models says the company is determined to power and accelerate the next computer era.
As computing continues to shift from retrieval-based generative models, NVIDIA focuses on tokens as the core currency of AI and its extended capabilities across cloud, enterprise and robotics platforms, providing a roadmap for the future of technology, with a great impact on global industries.
