AWS is changing the cost equation for enterprises weighing up where to put AI infrastructure for processing sensitive data.
Enterprises wanting to process sensitive data with AI face a dilemma: Pay more for on-premises AI infrastructure, or battle with data sovereignty and compliance barriers in the cloud. Now AWS is weighing in with another option: a fully managed, on-premises AI infrastructure offering it calls AI Factories.
Much like its existing AWS Outposts offering for classical compute workloads, AI Factories drops dedicated hardware and software directly into a customer’s data center, helping them run AI and agentic applications without violating data sovereignty rules.
“With this launch, we’re enabling customers to deploy dedicated AI infrastructure for AWS in their own data centers for exclusive use for them,” CEO Matt Garman said during his keynote speech at the AWS re:Invent customer conference.
AI Factories’ dedicated hardware and software stack includes Nvidia’s latest GPUs, AWS Trainium chips, high-performance networking, and software services such as SageMaker and Bedrock.
Garman positioned it as a “private AWS region” enabling enterprises to leverage their own data center space and power consumption capacity, giving them cloud-like elasticity on infrastructure they fully control — a growing priority as data sovereignty pressures mount.
The service will help enterprises facing AI adoption challenges complicated by data sovereignty and compliance, said HyperFRAME Research analyst Stephen Sopko.
“The AWS AI Factory seeks to resolve the tension between cloud-native innovation velocity and sovereign control. Historically, these objectives lived in opposition. CIOs faced an unsustainable dilemma: choose between on-premises security or public cloud cost and speed benefits,” he said. “This is arguably AWS’s most significant move in the sovereign AI landscape.”
On premises GPUs are already a thing
AI Factories isn’t the first attempt to put cloud-managed AI accelerators in customers’ data centers. Oracle introduced Nvidia processors to its Cloud@Customer managed on-premises offering in March, while Microsoft announced last month that it will add Nvidia processors to its Azure Local service. Google Distributed Cloud also includes a GPU offering, and even AWS offers lower-powered Nvidia processors in its AWS Outposts.
AWS’ AI Factories is also likely to square off against from a range of similar products, such as Nvidia’s AI Factory, Dell’s AI Factory stack, and HPE’s Private Cloud for AI — each tightly coupled with Nvidia GPUs, networking, or software, and all vying to become the default on-premises AI platform.
But, said Sopko, AWS will have an advantage over rivals due to its hardware-software integration and operational maturity: “The secret sauce is the software, not the infrastructure,” he said.
Omdia principal analyst Alexander Harrowell expects AWS’s AI Factories to combine the on-premises control of Outposts with the flexibility and ability to run a wider variety of services offered by AWS Local Zones, which puts small data centers close to large population centers to reduce service latency.
Sopko cautioned that enterprises are likely to face high commitment costs, drawing a parallel with Oracle’s OCI Dedicated Region, one of its Cloud@Customer offerings.
“Oracle’s comparable Dedicated Region requires a five-year commitment with an annual minimum spend starting around $1 million. Given the infrastructure investment AWS must make deploying hardware, networking, and management I would expect similar multi-year commitments,” Sopko said.
AWS hasn’t disclosed specific commitment terms.
