Next-gen NVIDIA GPUs
Blackwell, Blackwell Ultra and Vera Rubin on the Akash Supercloud, from single GPUs on demand to reserved rack-scale clusters.
Models marked “On the network” are listed by providers today. For everything else, tell us what you need and the team will get in touch about capacity.
See live GPU prices- Register interest
NVIDIA Vera Rubin
Rubin · next generation after Blackwell
NVIDIA's next platform pairs Rubin GPUs with HBM4 memory and the Vera CPU, built for frontier-scale training and inference. Register interest to plan Rubin capacity on Akash.
HBM4 - Request access
NVIDIA HGX B300
Blackwell Ultra
The most memory per GPU in the Blackwell family, for serving trillion-parameter MoE models and long-context inference on a single 8-GPU node.
288 GB HBM3e per GPU - Reserved clusters
NVIDIA GB300 NVL72
Grace Blackwell Ultra
Rack-scale Blackwell Ultra: 72 GPUs in one NVLink domain, for reasoning models and the largest training runs. Offered as reserved clusters.
72 GPUs · 36 Grace CPUs per rack - Request access
NVIDIA HGX B200
Blackwell
Blackwell with native FP4 for training and inference. An 8-GPU node brings 1.44 TB of GPU memory linked by 1.8 TB/s NVLink.
180 GB HBM3e per GPU - Reserved clusters
NVIDIA GB200 NVL72
Grace Blackwell
Liquid-cooled rack that links 72 Blackwell GPUs and 36 Grace CPUs as one system, for real-time LLM inference and large-scale training. Offered as reserved clusters.
72 GPUs · 36 Grace CPUs per rack