Amazon Web Services (AWS) has made its latest generation Trainium chips – Trainium3 – available via its UltraServers offering.

UltraServers were launched by AWS in December 2024. Comprised of four connected instances, they were initially offered with AWS' Trainium2 chips, with each UltraServer offering 64 chips.

Trainium3 is the company's first 3nm AI chip, and according to the company, the Trainium3 UltraServer offers up to 4.4x more compute performance, four times greater energy efficiency, and nearly four times more memory bandwidth than the previous generation.

In addition, the UltraServers use the new NeuronSwitch-v1, which delivers twice as much bandwidth within each UltraServer, while enhanced Neuron Fabric networking reduces communication delays between chips to under 10 microseconds.

Announced at the AWS Re:Invent 2025 event in Las Vegas, the Trainium3 UltraServers, now generally available, can scale up to 144 Trainium3 chips, offering 362 FP8 petaflops and four times lower latency for training large models.

Customers can also connect thousands of UltraServers to connect up to 1 million Trainium chips – around 10x as many as the previous generation.

According to AWS, customers that are already using the Trainium3 UltraServers include Anthropic, Karakuri, Metagenomics, Neto.ai, Ricoh, and Splashmusic.

Matt Garman, AWS' CEO said during his Keynote that AWS has thus far deployed more than 1 million Trainium chips, and that the company has done it in "record speed."

Speaking on this, Garman explained this rapid roll-out was because "we [AWS] control the whole process, the whole stack, we can optimize how we roll it out, and that allows us to move even faster."

He added: "In fact, we've been able to ramp the volumes of Trainium2 in our data center 4x faster than the next fastest AI chip we've ever ranked, and we're selling those as fast as we can make them. Trainium already represents a multi-billion-dollar business, and continues to grow really rapidly."

In addition to making Trainium3 available, AWS has teased the upcoming Trainium4 chips, which it says will bring "significant performance improvements across all dimensions."

Trainium4, AWS says, will have at least 3x the FP8 processing power and 4x more memory bandwidth than Trainium3. In addition, Trainium4 will be designed to support Nvidia NVLink Fusion chip interconnect technology.

Earlier this year, AWS made UltraServers based on the Nvidia GB200 NVL72 system generally available. In addition to launching the latest Trainium UltraServers, AWS announced during the Re:Invent conference that it was now offering the P6e-GB300 UltraServers, which feature the Nvidia GB300 NVL72 system, describing it as ideal for "AI inference at scale supporting trillion-parameter models with reasoning capabilities in production."