On 16 March 2026 NVIDIA announced that the Vera Rubin platform — Vera CPUs plus Rubin GPUs, NVLink 6, and related networking — was in full production. It says Vera Rubin NVL72 trains large mixture-of-experts models with a quarter of the GPUs versus Blackwell, and claims up to 10× inference throughput per watt.
Those are vendor numbers. What to watch is who actually ships a rack, at what power, and at what price per token. AWS, Google Cloud, Azure, Oracle, and several GPU clouds are named as buyers.
The story under the chips is still electricity, cooling, and who can afford the next factory. A faster rack does not make a model’s answer true.
OriginalNVIDIAinvestor.nvidia.comRead the full story