- AWS urges customers to change Nvidia to their cheapest training chip
- He says that his hardware offers the same yield with cost savings of 25 percent
- Amazon’s tone occurred when Nvidia showed her new hardware in GTC 2025
While Nvidia was organizing its annual GTC 2025 conference, showing new products such as IA DGX Spark and DGX Station AI supercomputers, Amazon was trying to convince its clients in the cloud that they could save money by moving away from the expensive Nvidia hardware and adopting the Aiazon AI chips themselves.
The information AWS’s statements launched at least one of their cloud customers to consider renting servers driven by the Amazon Training Chip, claiming that they could enjoy the same performance as NVIDIA H100, but with 25 percent of the cost.
Training is one of several internal chips that Amazon has developed (together with Graviton and Inferentia), built to train automatic learning models in the AWS cloud and offer a lower cost alternative to GPU -based systems. Amazon silicon does not intend to be a similar replacement for the most advanced products in NVIDIA, but it is not necessary.
Part of the conversation of AI
Amazon’s offer seems to be part of a broader change in the cloud market, where suppliers such as AWS and Google are developing their own chips and offering them to customers as a way of avoiding the cost and shortage of highly wanted NVIDIA GPUs.
“What AWS is doing is intelligent,” said Matt Kimball, vice president and main analyst for data calculation and storage in Moor Insights & Strategy. Network world. “That is to tell the world that there is a profitable alternative that is also a performance for the training needs of AI. It is inserting in the conversation of AI.”
The launch here, of course, is access. AWS gives customers the opportunity to experiment with the training and inference of workloads without having to wait months for a NVIDIA GPU or pay the best price for it.
While 25 percent savings definitely should not be sniffed, and something that will undoubtedly attract several AWS clients, there are obvious disadvantages for buyers to consider.
As Network world Notes: “Companies used to work with the architecture of Unified Compute de Nvidia (CUDA) should think about the cost of changing to a completely new platform as training. In addition, training is only available on AWS, so users can block.”
