AWS introduces AI factory model for governments and public-sector workloads

According to AWS, its AI factories have dedicated infrastructure with NVIDIA’s computing platform, as well as AWS Tranium chips, AWS AI services and AWS networking.

Matt Garman, AWS CEO.

Amazon Web Services (AWS) has outlined a new deployment model designed to help public-sector organisations build and operate artificial intelligence systems within their own facilities. The company plans to make “AI factories” available to agencies that require local control of data, infrastructure and compliance processes.

In this model, an AI factory refers to the combined hardware, software and operational framework needed to develop, train and run AI models at scale. AWS says the setup will draw on a mix of NVIDIA accelerators, its own Trainium chips, managed AI services and cloud-networking technologies. While customers supply the data-centre space, power and connectivity, AWS is responsible for installing, operating and integrating the environment.

The approach is aimed at entities that cannot move certain datasets to public cloud regions because of data-sovereignty or regulatory restrictions. By hosting the infrastructure on-premises, these organisations can train and run large models on proprietary information without relocating sensitive material.

The announcement follows a partnership signed in November between AWS and Humain, a Saudi public investment initiative focused on AI development. Under that agreement, the companies plan to deploy roughly 150,000 AI accelerators—including NVIDIA’s GB300 systems and AWS Trainium hardware—in a dedicated AI zone in Riyadh. The collaboration involves a projected US$5 billion investment covering infrastructure, cloud services and workforce training, with the intention of serving customers inside and outside the region.

Speaking at AWS re:Invent in Las Vegas, AWS CEO Matt Garman said government organisations in multiple countries had raised interest in a model that allows AI compute to operate inside their own data centres. He described the AI factories as functioning similarly to a private AWS region, giving customers access to AWS-managed AI hardware and services while using the physical capacity they already maintain.

According to AWS, the factories will support technologies such as Trainium UltraServers, NVIDIA GPUs, Amazon SageMaker for model development, and Amazon Bedrock for generative-AI workloads. Garman also noted that the latest models from Anthropic and Amazon’s own Bedrock platform are running on Trainium hardware, and that AWS has deployed more than one million Trainium chips to date.

AWS additionally announced the general availability of Trainium 3 UltraServers, built around a 3-nanometre AI chip. The servers are designed to provide higher compute density and memory bandwidth, and the largest configuration links 144 Trainium 3 chips through custom interconnects, delivering 362 petaflops of compute and more than 700TB/s of aggregate bandwidth.