Instances now generally available
Amazon Web Services (AWS) has made UltraServers based on the Nvidia GB200 NVL72 system generally available.
UltraServers were launched by AWS in December 2024. Comprised of four connected instances, they were initially offered with AWS' Tranium2 chips, with each UltraServer offering 64 chips.
The new UltraServer instance, dubbed Amazon EC2 P6e-GB200 UltraServers, has either 36 or 72 Nvidia Blackwell GPUs within one NVLink domain, and can offer up to 360 petaflops of FP8 compute and 13.4TB of total high bandwidth memory (HBM3e).
The UltraServers can further be connected with AWS Nitro System, scalable up to tens of thousands of GPUs.
According to AWS, the instance type is ideal for "the most compute and memory-intensive AI workloads" such as the training and inference of "frontier models" at the trillion parameter scale.
The instances are currently available in the Dallas Local Zone as an extension of the US East (Northern Virginia) region.
AWS recently cut costs significantly for instances with Nvidia H100 and H200 chips, up to 45 percent in some cases.
AWS made Nvidia B200 GPUs generally available in May 2025 and H100s in July 2023.
