AWS says Trainium2 cuts training costs about 40% versus Nvidia

Amazon Web Services’ claim that Trainium2 can reduce AI training costs by about 40% versus Nvidia is real, but the company’s published wording is more specific than some social-media reposts suggested. AWS says its Trn2 instances, powered by Trainium2 chips, deliver “30-40% better price performance” than GPU-based EC2 P5e and P5en instances, which are Nvidia-based offerings on AWS. The posts circulating on X on Tuesday, May 19, appear to be amplifying an AWS marketing line that has been on the company’s site since the Trainium2 launch in December 2024. AWS announced general availability for Trainium2-powered EC2 Trn2 instances and introduced Trn2 UltraServers at re:Invent on Dec. 3, 2024. (aws.amazon.com) ### What exactly did AWS say? AWS says on its Trn2 product page that “Trn2 instances offer 30-40% better price performance than GPU-based EC2 P5e and P5en instances.” The same language appears in the AWS News Blog and Amazon’s press release announcing Trainium2 availability. That phrasing matters because AWS is not claiming a blanket 40% advantage over all Nvidia hardware or all cloud GPU options. (aws.amazon.com) The published comparison is tied to two specific AWS instance families — P5e and P5en — rather than the entire Nvidia lineup. ### Which Nvidia systems is Trainium2 being measured against? AWS names EC2 P5e and P5en as the comparison set. (aws.amazon.com) Those are GPU-based AWS instances, and AWS’s wording frames the benefit as price performance, not simply lower hourly list price. That distinction leaves room for AWS to combine cost and throughput in one metric. In other words, the company is saying customers may get more training output per dollar on Trn2 than on those Nvidia-backed AWS instances, not necessarily that every Trn2 deployment will show a straight 40% lower bill. (aws.amazon.com) ### What is Trainium2 selling alongside the cost claim? (aws.amazon.com) AWS says Trn2 instances are built around 16 Trainium2 chips and target training and deploying models ranging from hundreds of billions to more than a trillion parameters. The company says a Trn2 instance offers up to 20.8 FP8 petaflops, 1.5 TB of HBM3 memory and 3.2 Tbps of EFAv3 networking. (aws.amazon.com) AWS also says Trn2 UltraServers connect 64 Trainium2 chips with its NeuronLink interconnect. On the product page, AWS says UltraServers are available now for demanding generative AI workloads; in the December launch materials, AWS said the systems were introduced as a new compute offering for larger-scale training and inference. ### Why are the May 19 posts getting attention now? (aws.amazon.com) The broader AI-chip discussion on X has focused on whether large cloud companies can reduce dependence on Nvidia by pushing their own accelerators. The Trainium2 posts fit that theme, but the underlying AWS claim is not new. It dates to the December 2024 launch and remains live on AWS’s Trainium and EC2 pages. (aws.amazon.com) AWS has paired that message with a larger build-out. In December, the company said it was working with Anthropic on Project Rainier, an EC2 UltraCluster containing hundreds of thousands of Trainium2 chips. ### What should readers treat as verified? The verified part is AWS’s own published statement: Trainium2-based Trn2 instances offer 30-40% better price performance than AWS EC2 P5e and P5en instances. (aws.amazon.com) The less certain part is any broader social-media shorthand that turns that into a universal “40% cheaper than Nvidia” claim without naming the benchmark, workload, or instance family. (press.aboutamazon.com) AWS’s current product pages are the clearest source for the wording, and they remain the place to check for any updated pricing, specifications, or comparison details. (aws.amazon.com)