Before a standing-room-only crowd, Huang introduced the Nvidia GB300 NVL72, the company's most advanced AI inference platform to date, alongside a suite of cutting-edge products spanning superchips, servers, robotics platforms, and quantum-ready systems. "AI is ushering in a new industrial revolution," Huang declared. "Every company will need to build or rent an AI factory."
The keynote marked a bold strategic repositioning of Nvidia as an AI infrastructure company, with the new Grace Blackwell GB300 platform at the center of its vision.
This liquid-cooled behemoth packs 72 Blackwell Ultra GPUs and 36 Grace CPUs into a single rack, delivering a 10x improvement in user response speed and 5x throughput over its predecessor Hopper, thanks to expanded HBM memory and doubled network connectivity.
The system is purpose-built for real-time inference workloads, including Agentic AI and Physical AI applications, which Huang emphasized as the next frontier toward AGI. With next-gen interconnects like NVLink Fusion and Quantum-X800 InfiniBand, Nvidia says the GB300 platform can drive a 5000% increase in model inference efficiency — effectively turbocharging the AI "factory" concept.
In addition to datacenter breakthroughs, Nvidia introduced the RTX PRO Server, built on the Blackwell-based RTX PRO 6000 GPU, aimed at enabling GPU-accelerated enterprise workloads. Partnering with global OEMs like Cisco, Dell, HPE, and Lenovo, Nvidia is deploying the RTX PRO Server as part of its Enterprise AI Factory framework.
Huang also revealed DGX Spark and DGX Station, AI-first personal computing systems powered by the GB300 Ultra desktop superchip, delivering up to 20 petaflops of AI performance and 784GB of memory. Shipping in July, the systems support high-speed 800Gb/s networking via ConnectX-8 SuperNIC.
Meanwhile, the HGX GB300 NVL72 supercomputing system enters mass production in Q3, boasting 72 GPUs, 288GB of HBM3E memory, and a 50% performance boost over the GB200 NVL72. Orders, Huang noted, are already flooding in.
One of the keynote's biggest moments came when Huang pulled back the curtain on Nvidia's next-gen AI supercomputer, which integrates 1.3 trillion transistors, two miles of copper wire, and a 130TB/s NVLink backbone. Built using a new TSMC-Nvidia COOS-L process, the system will weigh 1.8 tons and consist of 1.2 million components, assembled by partners including Foxconn, Wistron, Dell, and ASUS.
Nvidia also introduced the Grace CPU C1, optimized for edge, telecom, and cloud workloads, boasting 2x the energy efficiency of traditional CPUs — crucial for power-constrained AI inference at the edge.
In quantum computing, Nvidia unveiled the Global Quantum AI Technology Commercial R&D Center ( G-QuAT ) and ABCI-Q, the world's largest quantum AI supercomputer. Huang reversed his previous skepticism about quantum's timeline, now predicting all future supercomputers will integrate CPUs, GPUs, and QPUs for hybrid AI-quantum workloads.
Huang emphasized that Physical AI — the fusion of AI and robotics — is a trillion-dollar opportunity. Central to that vision is Newton, Nvidia's newly announced, GPU-accelerated physics engine developed with DeepMind and Disney Research. Set to open-source in July, Newton will power next-gen robot training within Nvidia's Isaac Sim 5.0 environment.
Nvidia also launched Isaac GR00T N1.5, its foundational model for humanoid robot reasoning, and Cosmos Reason, a world model trained on over 24,000 robot motion trajectories. New partners like Boston Dynamics, Agility Robotics, Fourier Intelligence, XPeng Robotics, and General Robotics are already leveraging the platform.
In collaboration with Hon Hai ( Foxconn ) , Nvidia will build an AI factory supercomputer equipped with 10,000 Blackwell GPUs. "These factories are the new engines of productivity," Huang said, predicting that digital agents and humanoid robots will offset a projected global labor shortfall of 30 – 50 million by 2030.
TSMC CEO C.C. Wei echoed that vision, stating: "By embracing Nvidia's infrastructure, we're accelerating breakthroughs in semiconductor innovation and beyond."
From AI-first PC systems to datacenter powerhouses, edge inference chips, robotics platforms, and quantum computing, Nvidia's Computex 2025 showcase underscores its transformation into an end-to-end AI platform company.
By integrating Agentic reasoning, Physical AI, and next-gen infrastructure, Huang has positioned Nvidia not just as a chipmaker, but as the engine behind the global AI revolution.
