AWS Trainium is the AI accelerator Amazon Web Services designs for training and running AI models, available to customers through Amazon EC2 instances. As of October 2026, AWS offers Trainium2 in Trn2 instances and UltraServers and Trainium3 in Trn3 UltraServers, and has announced Trainium4. The best-known model trained on Trainium is Anthropic’s Claude: Amazon says Anthropic uses the chips “to build, train, and deploy” its models.

On this page

What Trainium is for

AWS describes Trainium as “purpose-built AI chips to deliver high performance and the best economics for AI training and inference at scale.” Customers use the chips as EC2 capacity: Trn2 instances, Trn2 UltraServers and Trn3 UltraServers. Amazon also runs its own services on them; Amazon Bedrock is among the Trainium3 users Amazon names.

The unit of scale has grown with each generation. A single Trn2 instance holds 16 Trainium2 chips. A Trn2 UltraServer joins 64 of them with AWS’s NeuronLink interconnect, and a Trn3 UltraServer holds up to 144 Trainium3 chips on a switched fabric. More chips on one fast network let a model too big for a single machine be split across many chips with less time lost to communication between them.

Trainium2: Trn2 instances and UltraServers

AWS lists the trn2.48xlarge instance with 16 Trainium2 chips connected by NeuronLink, 1.5 TB of HBM3 with 46 TB/s of memory bandwidth, and 20.8 FP8 petaflops of compute, plus 192 vCPUs and 2 TB of host memory. The Trn2 UltraServer connects 64 Trainium2 chips for 6 TB of HBM, 185 TB/s of memory bandwidth and 83.2 FP8 petaflops.

AWS says Trainium2 supports FP32, TF32, BF16, FP16 and a configurable FP8 (cFP8) format, with 4x sparsity (16:4), stochastic rounding and dedicated collective engines.

Trainium3: Trn3 UltraServers

Amazon announced on December 2, 2025 that Trn3 UltraServers were generally available. Trainium3 is, in AWS’s words, its “first 3nm AWS AI chip.” Each chip carries 144 GB of HBM3e at 4.9 TB/s, which AWS says is 1.7x the memory bandwidth of Trainium2, and AWS says the chip delivers up to 2x the MXFP8 compute of Trainium2.

A Trn3 UltraServer scales to 144 Trainium3 chips with up to 362 MXFP8 petaflops, 20.7 TB of HBM3e and 706 TB/s of aggregate memory bandwidth. The chips are joined by NeuronSwitch-v1, an all-to-all fabric with 2 TB/s of bandwidth per chip, double the inter-chip bandwidth of Trn2 systems according to AWS. For larger jobs, AWS says Trn3 UltraServers are available in EC2 UltraClusters 3.0 that “scale to hundreds of thousands of chips.”

AWS’s comparisons with Trn2 UltraServers: up to 4.4x higher performance, more than 4x better energy efficiency and more than 5x higher output tokens per megawatt.

Trainium at a glance

Figures as AWS lists them. Compute formats differ between rows (FP8 for Trainium2, MXFP8 for Trainium3).

System Chips HBM Memory bandwidth Peak compute
Trn2 instance (trn2.48xlarge) 16 × Trainium2 1.5 TB HBM3 46 TB/s 20.8 PFLOPs FP8
Trn2 UltraServer 64 × Trainium2 6 TB 185 TB/s 83.2 PFLOPs FP8
Trainium3 chip 1 144 GB HBM3e 4.9 TB/s Up to 2x Trainium2 (MXFP8)
Trn3 UltraServer Up to 144 × Trainium3 20.7 TB HBM3e 706 TB/s Up to 362 PFLOPs MXFP8

Specs for Trainium and other accelerators are in our AI chip catalog.

Trainium4

In the same December 2, 2025 announcement, Amazon said Trainium4 is in development and is designed to deliver “at least 6x the processing performance (FP4), 3x the FP8 performance, and 4x more memory bandwidth” than Trainium3. Amazon also said Trainium4 is “designed to support NVIDIA NVLink Fusion high-speed chip interconnect technology,” which would let Trainium4 systems use Nvidia’s chip-to-chip interconnect.

Which models train on Trainium

Anthropic’s Claude. Project Rainier, which Amazon describes as one of the world’s largest AI compute clusters, links “nearly half a million Trainium2 chips” across multiple US data centers, including a site in St. Joseph County, Indiana. Amazon says Anthropic is “actively using Project Rainier to build and deploy” Claude, and in late 2025 said Claude was expected to be on more than 1 million Trainium2 chips, for training and inference, by the end of that year.

On April 20, 2026, Amazon announced an expanded agreement under which Anthropic “will secure up to 5 gigawatts (GW) of capacity to train and power their advanced AI models.” Amazon says the commitment covers Trainium2, Trainium3, Trainium4 and future generations, and that Anthropic will run its large language models on Trainium “for the next decade.” Projects of this size are listed in our AI data center tracker.

Other customers AWS names. AWS’s Trainium pages list Anthropic, Databricks, Decart, OpenAI, Ricoh, Splash Music and Uber among Trainium users, and name poolside, Itaú Unibanco and Brave among Trn2 customers. For Trainium3, Amazon names Anthropic, Karakuri, Metagenomi, NetoAI, Ricoh, Splash Music, Decart and Amazon Bedrock, and says these customers are “reducing their training costs by up to 50% compared to alternatives.” Decart, according to Amazon, gets “4x faster frame generation at half the cost of GPUs.” These are AWS’s and the customers’ own claims. Profiles of the labs are in the AI companies directory.

Sources