As AI models grow increasingly large, governments and large enterprises are placing greater emphasis on data sovereignty and compliance. At re:Invent 2025, AWS announced its new AI Factories service, which deploys AWS's AI infrastructure, including the latest NVIDIA accelerated computing platform and AWS's proprietary Trainium chips, directly into customers' existing data centers, helping them quickly build high-performance, compliant, and sovereign AI computing environments.
Addressing the pain points of self-built systems, operating like a "private AWS region".
AWS points out that for regulated industries and the public sector, building large-scale AI infrastructure in-house will face significant challenges in terms of capital investment and procurement cycles. The core concept of AWS's AI Factory service is to move AWS's complete AI stack (including high-speed networks, storage, security, and services such as Bedrock and SageMaker) directly into the customer's data center, with AWS responsible for operation.
This makes the environment function like a "Private AWS Region," where customers can utilize existing power and space while enjoying AWS's managed services and model access without having to negotiate licenses with multiple vendors, significantly reducing deployment time and meeting regulatory requirements for data localization.
Deepening the partnership with NVIDIA will involve the adoption of Grace Blackwell and the upcoming Vera Rubin.
At the hardware level, the collaboration between AWS and NVIDIA will be further deepened. AWS AI Factory services will integrate NVIDIA's full-stack AI software and accelerated computing platform, including support for the latest NVIDIA Grace Blackwell accelerated computing architecture, as well as upcoming offerings.NVIDIA Vera Rubin Computing Platform.
In addition, AWS's next-generation Trainium 4 chip will also support it.NVIDIA NVLink Fusion high-speed interconnect technologyThis means that, following Qualcomm, MediaTek, Intel, and other companies, AWS will also more closely integrate with NVIDIA's GPU-accelerated computing ecosystem through NVLink Fusion technology, thereby giving AWS customers greater flexibility in choosing AI-accelerated computing options.
Ian Buck, VP of Hyperscale and HPC at NVIDIA, said that large-scale AI computing requires a holistic approach. This collaboration with AWS will enable the delivery of these massive computing capabilities directly to customer environments, allowing organizations to focus on innovation rather than integration.
The P6e-GB300 UltraServer debuts, with initial deployments in Saudi Arabia.
Alongside the launch of its AI Factory service, AWS also announced the new Amazon EC2 P6e-GB300 UltraServers instance, which features...NVIDIA GB300 NVL72 systemBuilt specifically for large-scale AI inference operations, it supports trillion-parameter models with reasoning capabilities in production environments and is powered by AWS Nitro System, enabling seamless integration with services such as Amazon EKS.
The first application case for the P6e-GB300 UltraServer will be for the Saudi Arabian Public Investment Fund.DESIGNThe first "AI Zone" is planned to be established in Saudi Arabia. The zone will deploy up to 150000 AI chips (including NVIDIA GB300 GPUs), powered by AWS AI Factory service infrastructure, to meet the rapidly growing AI computing power demand locally and globally.
In addition to collaborating with AWS and NVIDIA, HUMAIN has previously announced partnerships with AMD, Qualcomm, and other industry players.AI computing chip supplyFurthermore, at Snapdragon Summit 202 this year, they even announced a collaboration with Qualcomm to launch...AIPCThis will drive the development of more user-friendly AI agents and also allow Qualcomm AI-accelerated computing solutions to be deployed in Saudi Arabia.Large-scale AI inference infrastructure.
In addition to providing Amazon EC2 P6e-GB300 UltraServers, AWS will continue to provide Amazon EC2 P6e-GB200 UltraServers with GB200 NV72 system, and will also continue to provide Amazon EC2 P6 with B300 and B200 systems.
