Nvidia’s Vera Rubin platform has entered full production and is now shipping to the world’s largest cloud operators. The new NVL72 rack-scale systems, each packing 72 Rubin GPUs and 36 Vera CPUs, represent a generational leap in AI inference efficiency.
For crypto investors, the more interesting subplot is happening off to the side. Bitcoin miners, sitting on massive power contracts and purpose-built data center facilities, are rapidly repositioning themselves as AI infrastructure providers, hosting Nvidia GPUs instead of (or alongside) their ASIC rigs.
What Rubin actually delivers
The numbers here are striking. Nvidia’s Rubin architecture boasts up to 10x the throughput for agentic AI workloads per unit of energy consumed, compared to the previous Blackwell generation. Inference costs drop by a factor of ten as well. And this isn’t coming from some massive increase in silicon complexity. The transistor count only ticked up slightly from its predecessor.
The systems scale from the base NVL72 configuration to setups exceeding 1,000 chips. Supermicro began delivering end-to-end Rubin-based solutions in June 2026, including scalable units supporting up to 1,152 GPUs. CoreWeave landed among the first customers to receive orders, followed by the usual suspects: Microsoft, Amazon, and Oracle.
