AWS Project Rainier Deploys 500,000 Trainium2 Chips

AWS has brought Project Rainier online, completing the deployment in under 12 months since announcing the initiative at its re:Invent conference in December 2024. The cluster contains nearly 500,000 Trainium2 chips distributed across data centres in the United States.

Anthropic is using the infrastructure to train and deploy Claude, its foundation model. AWS projects that Anthropic will scale to more than one million Trainium2 chips by the end of 2025 for training and inference operations. The deployment represents a 70% increase in AWS's AI computing infrastructure. Project Rainier provides Anthropic with more than five times the compute power the company used to train earlier versions of its models.

AWS Trainium2 chips deliver training performance

Annapurna Labs, AWS’ custom silicon division, designed the Trainium2 chip for training foundation models and large language models. A single Trainium2 chip can complete trillions of calculations per second. The EC2 Trn2 instances feature 16 Trainium2 chips and deliver 20.8 peak petaflops of compute performance. AWS states the instances offer 30 to 40% better price performance than current GPU-based EC2 instances.

The architecture uses Trn2 UltraServers, which combine four physical servers into one unit. Each UltraServer contains 64 Trainium2 chips interconnected through NeuronLink, a connection technology developed by AWS. The NeuronLink connections use blue cables to distinguish them from other networking infrastructure in the data centre.

Ron Diamant, Distinguished Engineer at AWS and Head Architect of Trainium, says: “Project Rainier is one of AWS's most ambitious undertakings to date. It’s a massive, one-of-its-kind infrastructure project that will usher in the next generation of AI models.”

When multiple UltraServers are connected through Elastic Fabric Adapter networking technology, identified by yellow cables, they form what AWS calls an UltraCluster. This networking operates both within individual data centre buildings and across separate facilities. Project Rainier spans multiple AWS data centre locations rather than being concentrated in a single site. The project takes its name from the 4,392-metre stratovolcano visible from Seattle on clear days.

Anthropic scales Claude operations in Indiana

Matt Garman, CEO of AWS, told CNBC that Anthropic is already running about 500,000 chips in Indiana. “And in fact, it’s going so well that they've actually doubled down on that order,” he says.

Mike Krieger, Chief Product Officer at Anthropic, told CNBC that AWS’ ability to deliver infrastructure at scale distinguishes the partnership. “These deals all sound great on paper,” he says. “But they only materialise when they’re actually racked and loaded and usable by the customer. And Amazon is incredible at that.”

Amazon has invested US$8bn in Anthropic since the start of 2024. The partnership includes technical collaboration, with Anthropic providing input on infrastructure design.

Project Rainier sites implement water efficiency measures

The data centres in St Joseph County, Indiana maximise the use of outside air for cooling. AWS reports a water usage efficiency of 0.15 litres of water per kilowatt-hour. The company states this represents a 40% improvement since 2021.

The electricity consumed by Amazon’s operations, including its data centres, was matched 100% by renewable energy resources in 2023. Amazon has been the largest corporate purchaser of renewable energy in the world for the past five years.

AWS announced it would roll out new data centre components that reduce mechanical energy consumption by up to 46% and reduce embodied carbon in concrete by 35%. The sites constructed to support Project Rainier include these upgrades. The company maintains a path to be net-zero carbon by 2040.

AWS prepares Trainium3 chip deployment at Project Rainier

Matt told CNBC that AWS is preparing to deploy Trainium3, the third generation of its AI training chip. “It gives better performance, it gives better latency characteristics, it gets better power consumption per flop,” he says. “That will be deployed inside of Indiana. It’ll be deployed in many of our other data centres all around the world.”

Trainium3 is set to launch in the next few months. The chip was developed in collaboration with Anthropic, according to CNBC, with the AI company providing direct input to improve training speed, reduce latency and enhance energy efficiency.

Ron says: “When we build our own devices, we get to optimise across the entire stack to really compress engineering time and the time to get to massive scale.”