AWS re:Invent 2025 kicks off in Las Vegas in about ten days, and as someone who’s attended (physically or virtually) since the early days, I’ve developed a pretty good filter for separating the signal from the noise. Every year, AWS announces hundreds of services and features. Most are incremental improvements. A handful genuinely change how we build systems. Here’s what I’m watching for this year and why it matters.
The AI Infrastructure Play#
If there’s one thing we can predict with certainty, it’s that AI and machine learning will dominate the keynotes. AWS has been aggressively expanding its AI infrastructure — from custom Trainium and Inferentia chips to Bedrock’s growing model marketplace. The question isn’t whether there will be AI announcements; it’s whether they’ll be genuinely useful or just keeping-up-with-Azure announcements.
What I’m specifically watching for is improvements to Amazon Bedrock. The model-agnostic approach is smart — letting customers choose between Anthropic’s Claude, Meta’s Llama, and other models without rewriting their integration code. But the developer experience still has rough edges. Model versioning is confusing, the pricing model is opaque compared to calling APIs directly, and the fine-tuning workflows feel bolted-on rather than integrated.
If AWS announces a more streamlined fine-tuning pipeline with better cost visibility and perhaps native RAG (Retrieval-Augmented Generation) support that doesn’t require stitching together five different services, that would be genuinely valuable. The current pattern of “connect S3 to Kendra to Bedrock to Lambda to API Gateway” for a simple RAG application is exactly the kind of AWS complexity that drives teams to simpler alternatives.
Serverless Evolution#
Lambda has been the poster child for serverless computing for years, but it’s showing its age in some areas. Cold start times, while improved, are still a pain point for latency-sensitive applications. The 15-minute execution limit constrains certain workloads. And the programming model — while simple — doesn’t adapt well to complex, stateful workflows.
I’m expecting announcements around AWS Step Functions and workflow orchestration. The pattern of using Step Functions to coordinate Lambda invocations works but is verbose and hard to debug. If AWS introduces something like a simplified workflow DSL or better visual debugging tools, that would address a real pain point.
I’m also curious whether we’ll see any movement on serverless containers. AWS Fargate has been around for years, but the gap between “truly serverless” Lambda and “managed containers” Fargate is still wide. Something that offers container flexibility with Lambda-like scaling and pricing — pay per request, scale to zero — would be compelling.
The Database Landscape#
AWS already has an absurd number of database offerings — Aurora, DynamoDB, Neptune, Timestream, QLDB, MemoryDB, DocumentDB, Keyspaces, and more. Every year I half-expect them to announce “Amazon DatabaseForYourSpecificUseCase.” But beyond the jokes, there are genuine innovations happening here.
What I’d like to see is better cross-database querying. In any non-trivial application, data lives in multiple stores — relational data in Aurora, session data in ElastiCache, search in OpenSearch, analytics in Redshift. Querying across these boundaries is painful. Aurora Zero-ETL integrations with Redshift was a step in the right direction, but we need this pattern to expand.
DynamoDB improvements are always worth watching too. It’s become the default choice for serverless backends, but operations like data modeling, migration, and cost optimization remain challenging. Better tooling around capacity planning and cost prediction would help teams avoid the bill shock that comes from getting DynamoDB access patterns wrong.
Networking and Multi-Region#
This is the unglamorous category that quietly determines whether your architecture actually works at scale. AWS networking has improved enormously over the years, but multi-region deployments are still harder than they should be. AWS Cloud WAN simplified some of the network topology management, but the complexity of running a truly global application — with data residency requirements, latency-based routing, and regional failover — remains daunting.
I’d love to see improvements to cross-region replication for more services, simplified multi-region API Gateway setups, and perhaps better tooling for testing failover scenarios. The gap between “we’re multi-region” on the architecture diagram and “we’ve actually tested regional failover” is enormous in most organizations.
What I’m Skeptical About#
Every re:Invent brings announcements that sound impressive on stage but turn out to be limited preview features available in us-east-1 only, with pricing that makes them impractical for most workloads. I’m skeptical of any announcement that:
- Requires “contact sales for pricing” — it’s going to be expensive
- Is available in “preview” in one region — you won’t use it in production for a year
- Involves a new proprietary query language or SDK — the ecosystem won’t be there
- Claims to “simplify” something by adding another service to the stack
My Take#
The most impactful AWS announcements are rarely the flashiest. Lambda was a relatively quiet launch that changed how we build applications. The same goes for CDK, which didn’t get a keynote moment but fundamentally improved infrastructure-as-code on AWS. I’ll be watching for the quiet announcements that solve real problems, not the keynote demos that solve theoretical ones.
If you’re attending re:Invent, my advice is the same as every year: skip the keynotes (watch them on YouTube later), go to the chalk talks and builder sessions, and spend time in the hallway track talking to other practitioners. That’s where the real value is.
I’ll do a proper deep-dive on the actual announcements next month. For now, let’s see what AWS has in store.