NVIDIA GTC 2025: CEO Jensen Huang Unveils Next-Gen Blackwell Ultra B300 GPUs, Rubin and More
During the highly anticipated NVIDIA GTC 2025 keynote, CEO Jensen Huang revealed groundbreaking new technologies designed to accelerate the future of artificial intelligence and data-center computing. The announcements included details of the new Blackwell Ultra B300 GPU series, future AI computing platforms Rubin and Feynman, and powerful desktop AI systems DGX Spark and DGX Station.
Blackwell Ultra B300 GPUs: Unprecedented AI Computing Power
The centerpiece of the keynote was NVIDIA’s latest GPU innovation, the Blackwell Ultra B300 series. These GPUs are significant upgrades to NVIDIA's current Blackwell architecture, delivering unmatched performance specifically optimized for AI workloads. Featuring up to 288 GB of advanced HBM3e memory, the B300 series will significantly enhance data throughput, computational efficiency, and memory capacity—paving the way for breakthroughs in large-scale AI model training and inference tasks.
NVIDIA’s Ambitious AI Roadmap: Rubin and Feynman
Jensen Huang also provided an insightful look into NVIDIA’s next-generation AI roadmap, featuring two major upcoming architectures: Rubin and Feynman.
Vera Rubin NVL144 (Launching in late 2026)
The Vera Rubin platform, launching in the second half of 2026, will deliver groundbreaking AI performance through:
576 Rubin GPUs with dual reticle-sized GPU chips.
Up to 3.6 Exaflops of FP4 inference power, making it ideal for demanding AI workloads.
75 TB of ultra-fast HBM4 memory, offering unprecedented data access speeds.
Enhanced NVLink and networking capabilities for optimized large-scale deployments.
Rubin Ultra NVL576 (Arriving late 2027)
Following Vera Rubin, NVIDIA will introduce the Rubin Ultra NVL576 platform, scaling performance even further:
Up to 15 Exaflops of FP4 inference capability.
1 TB per GPU totaling 365 TB of next-generation HBM4e memory.
Upgraded NVLink interconnect technology, offering exceptional speed and scalability for future AI workloads.
Feynman GPUs (Slated for 2028)
Looking further ahead, NVIDIA confirmed the future GPU architecture codenamed Feynman, named after celebrated physicist Richard Feynman. Scheduled for 2028, Feynman GPUs will leverage next-generation memory technology (HBM4e or HBM5), enhancing both computational power and energy efficiency, and further demonstrating NVIDIA's long-term commitment to pushing the boundaries of AI and scientific computing.
DGX Spark and DGX Station: Enterprise-Level AI Performance on Desktops
Additionally, NVIDIA unveiled two revolutionary desktop solutions designed to bring high-performance AI computing directly to developers and researchers:
DGX Spark
The world’s smallest AI supercomputer featuring the new GB10 Grace Blackwell Superchip.
Capable of delivering up to 1,000 trillion operations per second for rapid local AI prototyping and inference.
Ideal for developers, data scientists, robotics engineers, and students, allowing intensive workloads to be executed right at their desks.
DGX Station
A desktop supercomputing powerhouse built upon NVIDIA’s GB300 Grace Blackwell Ultra Desktop Superchip.
Features an extraordinary 784 GB of coherent memory, optimized for training massive AI models.
Equipped with the NVIDIA ConnectX-8 SuperNIC, enabling networking speeds up to 800Gb/s for seamless integration into high-speed AI clusters and workflows.
Manufacturing partners such as ASUS, Dell, HP, and Lenovo will soon offer these advanced desktop systems, providing unprecedented AI performance in desktop environments.
Watch the Full NVIDIA GTC 2025 Keynote:
Experience Jensen Huang’s visionary keynote presentation in full detail here.
What do you think about NVIDIA's bold AI advancements announced at GTC 2025? Are these new products shaping the future of AI as you imagined? Let us know in the comments below!
