December 2025
[Andy Bold]
The annual AWS re:Invent took place in Las Vegas last week, accompanied by the usual tsunami of announcements on new and improved services.
If the last decade of cloud computing was defined by migrating infrastructure, and the last two years by experimenting with generative AI, then 2025 marks a decisive pivot. AWS CEO Matt Garman thinks we have reached an inflection point comparable to the advent of the Internet itself.
Even if that is the case, there were still plenty of non-AI announcements across Analytics, Compute, Security, and Storage. Here are the announcements that I found most impactful across the board.
Breaking the “AI Tax”
Several announcements directly addressed the costs and environment impacts of building and using AI solutions.
Trainium3 Ultraservers are available, delivering up to 4.4x more compute performance, 4x greater energy efficiency, and almost 4x more memory bandwidth than Trainium2. In practice, this means faster AI development with lower costs and a smaller environmental footprint.
For most companies this means simpler infrastructure for Retrieval-Augmented Generation (RAG) use cases, which is what we see most often at esynergy.
And with Amazon S3 Vectors, AWS has directly integrated vector indexing into S3. This serverless approach eliminates the need for expensive, dedicated vector databases. For most companies today this means that a less complex infrastructure is required for Retrieval-Augmented Generation (RAG) use cases. This is the use case that we see the most at esynergy, and we are looking forward to exploring the benefits to our customers.
It also simplifies Semantic Search across data stored in S3, which is another hot topic in our conversations with clients.
Taken together, cheaper compute and simpler data architectures go a long way to reducing the “AI tax” that has held back some experiments this year.
Amazon also signalled a lot of confidence in the Trainium platform when they shared that they expect similar improvements with Trainium4 in Q4 next year.
AI data sovereignty
One challenge with cloud-hosted AI solutions is that some businesses cannot make use of these solutions with their most sensitive and useful data due to regulatory, compliance, sovereignty, and privacy concerns.
AWS announced that AI Factories are available for businesses that have this constraint. With AI Factories, the customer provides hosting space for Trainium servers within their own data centres. Rather than invest CapEx in large infrastructure deployments, businesses can make use of this fully-managed service while still taking advantage of other services such as SageMaker and Bedrock.
With AI Factories, customers host Trainium servers in their own data centres. Rather than investing in large infrastructure deployments, they can use this fully managed service while still taking advantage of SageMaker, Bedrock, and other AWS AI services.
More capable AI models with Nova Forge
For organisations already running in AWS, the Nova family rounds out this picture on the model side.
Many of us are familiar with the model of buying an enterprise agreement from an AI supplier and integrating their models with our data, with guarantees about data use and privacy.
One of the challenges with this approach, especially if you have large data sets, is that fine tuning of the base models and integrations can lead to “catastrophic forgetting,” where the model learns new domain data but loses its general reasoning or safety alignment.
With Nova Forge AWS customers can use pre-trained model checkpoints that do not suffer from this problem, and blend their proprietary data with Nova’s own curated datasets.
It will be interesting to see how quickly the main AI providers will catch up with this.
AWS Clean Rooms
If Nova Forge tackles how you safely train and adapt models on your data, AWS Clean Rooms tackles how you safely create and share test data built from that same source.
Test data has been a perpetual thorn in my side for many, many years. There are various problems with test data sets.
- Developers want data to be as close to Production-like as possible, sometimes even demanding Production data to develop against. Notwithstanding regulatory concerns and restrictions in doing this, it is never a good idea.
- Database schemas evolve over time, and test data sets need to evolve with them. Version controlling is a chore that few teams are keen to take onboard, and test data versions often lag behind current Production state, meaning that Developers are building and testing against outdated schemas.
- Synthetic data generation tools have been available for a long time, but they often come with heavy costs, from both licensing and administrative overhead perspectives.
AWS Clean Rooms can generate synthetic datasets that retain the statistical patterns of the original sensitive data, but contains no actual real or identifiable data. It uses differential privacy techniques to guarantee protection against re-identification. And not only that, but it is possible for businesses to collaborate on using the synthesised data.
This could be particularly useful in Financial Services to collaborate on anti-fraud, KYC, and AML techniques, in healthcare to support public health projects without violating GDPR, and in advertising to analyse customer behaviour across partners without compromising privacy.
For teams that have always struggled to balance realistic test data with privacy and compliance, this feels like a meaningful step forward.
The Agentic Enterprise
AWS also introduced Frontier Agents, pre-packaged autonomous agents designed to perform specific high-value job functions. These are not just tools, they are digital workers intended to take on some of the work that currently adds load and stress to engineering teams.
Some of the more interesting ones are:
Amazon Kiro
Kiro aims to automate the bulk of the software delivery cycle. You describe the intended outcome, and it turns this into a specification, proposes an architecture, then generates the code and tests. The developer is engaged throughout the process, and can change and optimise the output from Kiro before it moves to the next stage.
AWS claims notable productivity gains from early adopters, such as senior engineers reporting that they have done more coding in six months with Kiro than they had in the previous three years.
The Kiro workflow aligns with esynergy’s internal investigations into spec-driven development, which reduces manual overhead for documentation, test creation, and boilerplate, and lets developers focus on design quality and delivery outcomes.
It will be interesting to see how Kiro compares to similar spec-driven frameworks from other services.
AWS DevOps Agent
As somebody who has been on call when `us-east-1` fails, I know that Incident Response is can be stressful. And when systems are becoming ever more complex and domain-driven, identifying root cause for a problem can be challenging, especially when under pressure to help fix Production services.
The AWS DevOps Agent has been released in Preview. It autonomously correlates signals from CloudWatch, and third party services such as Datadog, and code repositories to identify root causes. It can create mitigation plans, and update tickets in service management tools and send messages to Slack channels while it is investigating.
And it will build up a knowledge base by analysing past incidents to recommend proactive service improvements. Commonwealth Bank reported that using this agent reduced their Mean Time To Respond to Production incidents from hours to 15 minutes. This has a major impact on their service reliability, and the service that they provide to their customers.
AWS Security Agent
Penetration testing and security code reviews can be time consuming and delay time to market, while Security Engineers face an ever increasing volume of demands on their time.
Security audits are a quarterly or annual affair, leaving a lot of space for problems to creep into Production.
Security Agent continuously carries out code reviews, penetration testing, and vulnerability assessments, turning security into a continuous process that provides immediate feedback without slowing development.
Individual solutions for these problems do already exist, but using them can lead to tool sprawl, and an expensive, fragmented, and complex security tooling ecosystem.
AWS Transform for application migration and modernisation
Many organisations run years-old codebases and infrastructure that are hard to evolve. Modernisation is essential but often loses out to feature work, so technical debt accumulates and absorbs a sizeable share of the IT budget when it is finally addressed.
AWS Transform automates much of this effort. It can modernise .NET, Node.js, Java, and Python applications, updating dependencies, configuration, and APIs.
It supports migrations from x86 to Graviton for better price-performance, from on-premise to AWS as a first step to becoming cloud native, from CloudFormation to Terraform, and more.
Historically, at esynergy we would usually assemble small teams of experts to deliver similar modernisation work. With Transform we see an opportunity to accelerate these outcomes and with fewer experts needed, unlocking fast value for customers, including those who do not yet host on AWS.
Combined with the new Database Savings Plans, this removes some of the financial friction that has held back modernisation work.
Database Savings Plans
I saved the best to last.
It is finally possible to apply Savings Plans to managed databases. Previously, cost savings were limited to inflexible Reserved Instances with savings based on long term commitment to a certain database solution and instance size. This could lead overspend when the environment needed to change, and you were left paying for reservations that were no longer required.
Savings Plans will automatically migrate savings with you. So if you want to migrate from Provisioned to Serverless, or Oracle to PostgreSQL, the Savings Plan will move too.
This removes a constraint on modernisation, where initiatives have been held back by concerns about wasting money on a pre-paid reservation.
Conclusion
AWS announced over 40 improvements and new services last week. I have focused on the things that I found most interesting, but you should review the full list to see what applies most to you.
At esynergy we are positive about the benefits that AI tools can bring to supporting people in engineering. We do not see these tools as a replacement for people, but as a replacement for low value, high effort “grunt work” that people do not like to do, or that, with capabilities like Incident Response, can bring them stress and anxiety. The right tools free up time for engineers to focus on higher value work with more impact to the business, and based on the announcements at re:Invent, it seems that AWS feels the same.
Was Matt Garman right when he said that we have reached an inflection point as big as the internet itself? I don’t know if I would go that far. But there is certainly a lot of energy around people and businesses looking at these new tools and trying to understand how they can help, not necessarily replace, people.
The interesting work now is helping teams turn that energy into practical, safe, and sustainable changes in how they build and run systems.
