May 22, 2025 – At Computex 2025, NVIDIA CEO Jensen Huang delivered a groundbreaking keynote, unveiling the next evolution of AI computing with the Grace Blackwell GB300 system and the DGX Spark, alongside the revolutionary NVLink Fusion technology. These innovations mark a seismic shift in computing, transforming entire data centers into singular, AI-driven supercomputers and paving the way for a new era of “AI factories” that promise to reshape industries from enterprise IT to robotics.
The Power of Grace Blackwell: A Leap Beyond Moore’s Law
NVIDIA’s Grace Blackwell GB300 is a testament to the company’s relentless pursuit of computational excellence, achieving what Huang describes as “extreme Moore’s Law.” A single GBdeleted
System: * The user input is incomplete, as it ends abruptly with “achieving what Huang describes as ‘extreme Moore’s Law.’ A single”. I will assume the intended continuation is about the performance leap of the Grace Blackwell system, based on the provided context, and complete the response accordingly. If you have a specific continuation or additional details, please clarify, and I can adjust the response. *
The Power of Grace Blackwell: A Leap Beyond Moore’s Law
NVIDIA’s Grace Blackwell GB300 is a testament to the company’s relentless pursuit of computational excellence, achieving what Huang describes as “extreme Moore’s Law.” A single Grace Blackwell node delivers 40 petaflops of performance, equivalent to the Sierra supercomputer from 2018, which required 18,000 Volta GPUs. This represents a staggering 4,000x performance increase in just six years. The GB300 offers 1.5x more inference performance, 1.5x more HBM memory, and 2x more networking bandwidth compared to its predecessor, all within the same physical footprint, thanks to 100% liquid cooling and advanced chip design.
The secret to this scalability lies in NVIDIA’s NVLink technology, featuring a 7.2 terabytes per second switch and an NVLink spine with 130 terabytes per second of bandwidth—surpassing the peak traffic of the entire internet. This allows 72 GPUs to communicate simultaneously across a single rack, effectively turning the rack into a single, massive motherboard. Huang emphasized that this “scale-up” approach overcomes the limitations of semiconductor physics, enabling unprecedented computational power.
Why It Matters: Grace Blackwell enables inference-time scaling for reasoning AI, allowing systems to “think” by generating and iterating thoughts before producing answers. This is critical for advanced AI applications like real-time decision-making and complex problem-solving.
DGX Spark and DGX Station: AI Supercomputers for All
NVIDIA introduced the DGX Spark, a compact, 1-petaflop AI supercomputer with 128GB of HBM memory, designed for AI-native developers, researchers, and students. Available soon through partners like Dell, HPE, Asus, MSI, Gigabyte, and Lenovo, this desk-side system offers the same programming model as NVIDIA’s largest AI factories, making it ideal for prototyping and early development. Huang highlighted its evolution from the 2016 DGX-1, which had similar performance but was significantly larger and less efficient.
For those needing more power, the DGX Station pushes the limits of a standard wall socket, capable of running a 1-trillion-parameter AI model—far surpassing models like Llama 70B. Both systems democratize access to AI supercomputing, bringing enterprise-grade capabilities to individual users.
Why It Matters: These systems make high-performance AI accessible outside traditional data centers, empowering smaller teams and individuals to innovate in AI development.
NVLink Fusion: Customizable AI Infrastructure
NVIDIA’s NVLink Fusion introduces a new era of flexibility, allowing semi-custom AI infrastructure by integrating NVIDIA’s ecosystem with custom ASICs. The NVLink chiplet connects custom CPUs or accelerators to NVIDIA’s Blackwell and upcoming Rubin chips, enabling seamless integration into large-scale AI supercomputers. This open architecture ensures that companies can tailor their AI systems while leveraging NVIDIA’s high-performance networking and computing technologies.
Why It Matters: NVLink Fusion breaks down barriers, allowing diverse industries to build bespoke AI solutions within NVIDIA’s robust ecosystem, accelerating innovation across sectors.
AI Factories and the Future of Robotics
Huang described modern data centers as “AI factories,” with NVIDIA’s systems powering massive projects like the XAI Colossus factory and Oracle Cloud’s Stargate, a 1-gigawatt, 4-million-square-foot facility. These factories, costing $40-50 billion in computing alone, are designed for scale-out operations, connecting multiple systems for unparalleled performance.
In robotics, NVIDIA’s Isaac Groot platform and Jetson Thor processor enable AI-driven robotic systems, from self-driving cars to humanoid robots. The NVIDIA Isaac operating system handles neural network processing and sensor pipelines, while Omniverse provides a simulation environment for training robots. Huang emphasized the importance of synthetic data generation and human demonstration amplified by AI to teach robots complex skills, addressing labor shortages with versatile, scalable humanoid robots.
Why It Matters: The rise of AI factories and robotics represents a multi-trillion-dollar industry, with NVIDIA’s technologies enabling scalable, real-world applications in manufacturing, logistics, and beyond.
Digital Twins and Enterprise AI
NVIDIA’s Omniverse platform is central to creating digital twins—virtual replicas of factories, equipment, and robots. Companies like TSMC are using Omniverse to design digital twins for next-generation fabs, part of a $5 trillion global investment in new plants over the next three years. The RTX Pro Enterprise and Omniverse server, powered by the new Blackwell RTX Pro 6000 GPUs and CX8 networking chip, supports enterprise AI agents across text, graphics, and video modalities, ensuring compatibility with existing applications.
Why It Matters: Digital twins optimize plant design and robotic operations, reducing costs and accelerating deployment in the global re-industrialization wave.
Tags: NVIDIA, Grace Blackwell, DGX Spark, NVLink Fusion, AI Supercomputer, Robotics, Digital Twin, Omniverse, Computex 2025, AI Factory, Jetson Thor, Enterprise AI, Synthetic Data