NVIDIA's new Rubin platform promises to supercharge AI workloads
NVIDIA just introduced its Vera Rubin platform at GTC 2026, built to power next-gen AI factories.
With a set of new chips, including Rubin GPUs and Vera CPUs, it promises 10 times better efficiency for AI tasks and can train big models using far fewer GPUs than before.
Basically, it's a serious performance leap over the Blackwell platform.
NVL72 rack packs 72 Rubin GPUs
The NVL72 rack packs 72 Rubin GPUs (each with 288GB of fast HBM4 memory), providing 50 PFLOPS of NVFP4 compute per GPU.
There are also racks with 256 Vera CPUs, speedy LPDDR5X memory, and ultra-fast NVLink connections, meaning tons of muscle for heavy-duty AI work.
Chips are already in full production
Chips are already in full production; rack- and system-level offerings are slated to be available in the second half of 2026, with OEMs such as Dell Technologies, HPE, Lenovo, Supermicro, ASUS and Quanta Cloud Technology (QCT) rolling out servers.
The platform will ship with new networking tech like ConnectX-9 SuperNICs and BlueField-4 DPUs, so if you're into building or running giant AI systems, this is one to watch.