AWS launches AI factories for enterprise data centers

Amazon Web Services (AWS) has launched AWS AI Factories, an offering that lets organizations deploy dedicated AWS AI infrastructure within their own data centers.

The service combines Nvidia accelerated computing, Trainium chips, AWS AI services, and high-speed networking, with AWS handling deployment and management.

AWS said this approach is aimed at enterprises and public sector bodies needing to meet regulatory and data sovereignty requirements while accelerating AI rollout.

Customers use their existing data center space, network, and power, while AWS provides managed access to compute, storage, and AI model services.

🔗 Source: Amazon

🧠 Food for thought

Implications, context, and why it matters.

AWS AI Factories launch with on-premises deployments; pricing and commercial terms not disclosed

AWS AI Factories need customer data center space and power 1. Pricing and minimum footprint remain undisclosed. Contract terms and delivery timing are missing. The stack links to Amazon Bedrock (a managed service for accessing foundation models) and Amazon SageMaker (a managed machine learning platform) 1. Bedrock footnotes list a 75% premium for many models in the Priority tier 2. AWS has not posted AI Factory pricing.
NVIDIA GB200 NVL72 systems (a rack-scale AI platform built around NVIDIA Grace CPUs paired with Blackwell GPUs) draw about 120 kW per rack 3. They need liquid cooling with manifolds and cold plates (plumbing that circulates coolant directly to chips) 3. Typical CPU racks use near 12 kW 4. Many sites cannot handle that density 5. Retrofits for liquid cooling need care to keep safety and reliability 5.

Data center firms can win retrofit work from liquid cooling needs

Colocation providers (third-party data center operators that lease space, power, cooling) can bundle power upgrades and liquid cooling. Integrators can do the same. Most facilities lack 120 kW per rack 6 and the liquid systems needed for GB200-class gear 3. CoreWeave (a GPU-focused cloud provider) plans to make liquid cooling standard in most new data centers in 2025 5. Enterprise IT teams will lean on Data Center Infrastructure Management (DCIM) systems for power, environment, and capacity tracking 7.