Issue # 144: November 28, 2025, to December 12, 2025

Hey Reader.

Welcome to Issue # 144 of AWS Graviton Weekly, which will be focused on sharing everything that happened in the past week related to AWS Silicon: from November 28, 2025, to December 12, 2025.

After some crazy weeks, I let the dust settle, and now I will share the most interesting announcements, talks, and resources from the past AWS re:Invent 2025.

Every year, thousands of people from a lot of corners of the world gather together in Las Vegas to get the most exciting news from the AWS crew and from AWS partners in one single place.

I waited for the dust to settle because there are a lot of things to review and discuss, but my job here is to give you a condensed version of all that.

So, enjoy, and if you went to the event, share your thoughts in a reply. Love to hear about your experiences.

Recommendation of the week: CAST AI: 2025 GPU Price Report

Discover the availability and pricing of A100 and H100 GPU-based compute instances across AWS, Azure, and Google Cloud Platform.

Download the report

AWS announces Graviton5

For me, this was the most exciting announcement here: Graviton5. Why? The features:

192 cores
5x larger L3 cache
Deliver up to 25% higher performance than the previous generation
Graviton5 introduces the Nitro Isolation Engine as an enhancement to the Nitro System, harnessing formal verification to provide mathematical certainty that your workloads are isolated from each other and AWS operators

I know for sure that this new generation will bring more Enterprise customers to AWS, especially for high-intensive operations. We will continue monitoring the use cases for its adoption, so stay tuned.

Graviton5 is available in the M9g instances.

Trainium3 UltraServers are out

This was another announcement that could be very beneficial for companies doing highly intensive AI work. Why? Just read what they shared in the announcement posts:

Trn3 UltraServers scale up to 144 Trainium3 chips, delivering up to 362 FP8 PFLOPs with 4x lower latency to train larger models faster and serve inference at scale.

Crazy scale, right? And another aspect that surprised for good was the focus on decreasing costs as well:

Customers are already seeing significant value from Trainium, with companies like Anthropic, Karakuri, Metagenomi, NetoAI, Ricoh, and Splash Music reducing their training costs by up to 50% compared to alternatives. Amazon Bedrock, AWS's managed service for foundation models, is already serving production workloads on Trainium3, demonstrating the chip's readiness for enterprise-scale deployment.

If you have work in this kind of system, you know for sure that this stuff could get very expensive, and seeing AWS deeply thinking about being more efficient with AI ops is very refreshing.

Amazon EC2 X8aedz instances powered by 5th Gen AMD EPYC processors

AWS and AMD kept collaborating on many fronts, but this is one of my favorite ones. The AMD EPYC series has been a huge success for AMD, and AWS knows the power behind these chips, so this new generation had to be part of the AWS ecosystem.

The specs are very, very interesting for the instances