Nvidia CEO Jensen Huang announced during his keynote speech at CES 2026 that Vera Rubin is officially in full production. Vera Rubin NVL72 is Nvidia's next-generation AI-focused datacenter platform, and will feature up to 5x improved performance over Blackwell-equivalent products in some workloads.
Nvidia already showed off Vera Rubin at GTC 2025, but what's new is the fact that the Vera Rubin platform is now in production. Vera Rubin NVL72 takes advantage of an 88 Olympus core ARM-based CPU with "Spatial Multi-Threading," 128GB of GDDR7 memory, and a Rubin-based GPU with 288GB of HBM4 memory.
According to Huang, Vera's new multi-threading functionality enables each thread to have the full throughput of a single core, giving the chip the same processing capacity as 176 cores.
Accompanying these new chips are Nvidia's BlueField-4 DPU for offloading storage and security away from the CPU and GPU, Spectrum-6 Photonics Ethernet, and Quantum-CX9 InfiniBand NICs capable of up to 1.6 Tb/s. All of these chips are interlinked through Nvidia's high-speed 18TB/s NVLink interconnect.
As for the GPU — Nvidia claims its Rubin GPU is capable of 50 PFLOPs of NVFP4 inference, and 35 PFLOPs of NVFP4 training performance. These figures represent 5x and 3.5x the performance of Blackwell — all while increasing the GPU's transistor count by just 1.6x.
Nvidia has also made huge changes to Vera Rubin NVL72's chassis compared to previous generations — it's made the entire system fanless, tubeless, and cableless; the entire system is now 100% cooled with liquid cooling. Jensen claimed that installation times have gone from two hours on Blackwell-based counterparts down to just five minutes for Vera Rubin.
For more details, check out our main coverage of the Vera Rubin NVL72 platform. The new system will come out in the second half of 2026.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
