AWS UltraServers launch with NVIDIA Blackwell power
Powering the Next Generation of Cloud Computing
AWS has unveiled its UltraServers, a new cloud server platform built for AI and high-performance workloads. These machines use NVIDIA’s Grace Blackwell architecture to deliver faster training, lower latency, and greater scalability.
Each UltraServer features up to 72 NVIDIA Blackwell GPUs connected via NVLink. Backed by Grace CPUs, the system reaches 360 FP8 petaflops and over 13 TB of HBM3e memory. This setup delivers cloud performance that rivals supercomputers.
Built for Scale and Flexibility
The UltraServers integrate directly with the AWS Nitro System. This allows secure virtualization, strong performance isolation, and better control. As a result, businesses can scale workloads dynamically without overprovisioning resources.
Thanks to elastic support, enterprises can expand or reduce capacity based on demand. This flexibility reduces costs while improving efficiency across AI pipelines, data modeling, and inference.
Ideal for AI and Scientific Workloads
AWS is targeting sectors that demand serious compute power. These include autonomous systems, financial modeling, life sciences, and real-time analytics. Blackwell GPUs are built for generative AI, LLMs, and diffusion models, making them ideal for modern AI development.
In addition, UltraServers support hybrid architectures. Companies can run part of their workflows on-premise and burst into AWS when needed. This makes adoption smoother and more cost-effective.
The Future of Server Infrastructure
With the launch of UltraServers, AWS continues to push cloud infrastructure forward. The move reflects a larger trend: GPU-powered systems are now the foundation for enterprise-scale computing.
UltraServers give development teams the speed and power to run next-gen applications—without investing in their own hardware. They are now available in select AWS regions and will expand globally.
By offering this level of performance through the cloud, AWS is setting a new standard. These systems empower teams to build, train, and deploy faster than ever before.
Source: datacenterdynamics