AWS 2025: A Year of Agentic AI, Custom Chips, and Multicloud Bridges

Source: Dev.to

2025 AWS Year‑in‑Review

I’ve been tracking AWS releases all year, and honestly, 2025 felt different. Not just “here’s another managed service,” but fundamentally “we’re changing how you build software.” Between re:Invent and the steady stream of updates throughout the year, there’s a lot to unpack.

This article is my attempt to summarise the key announcements across every major category. I won’t cover everything—AWS announced hundreds of updates—but I’ll hit the ones that actually matter for most developers and architects.

Three Themes for 2025

Common thread: AWS is positioning itself not just as infrastructure, but as the platform where AI agents live and operate. Whether you buy into that vision or not, it’s clearly where they’re headed.

Global Footprint in 2025

New Regions Launched

• Mexico (Central) – AWS’s first region in Mexico, launched January 2025.
  • Commitment: $5 B over 15 years.
• Thailand (Asia Pacific) – Launched January 2025 with three Availability Zones (AZs).

Additional 2025 Launches

Coming Soon

• Saudi Arabia – Expected 2026.
• AWS European Sovereign Cloud – Launching Dec 2025 in Brandenburg, Germany with €7.8 B investment.

Note: As of Dec 2025, AWS operates 120 Availability Zones across 38 geographic regions. Remember—you don’t pay for an AWS account, only for the resources you run. Regional availability still matters for latency and data‑residency requirements.

Compute – The Big Wins

Graviton 5 Processors

• EC2 M9g instances: up to 25 % higher performance than the previous generation while using less energy.
• If you’re not already on Graviton for compatible workloads, the price‑performance gap just got wider.

Trainium 3 UltraServers

• Host up to 144 Trainium 3 chips per server.
• Deliver up to 362 MXFP8 PFLOPs of compute.
• Claimed 40 % more energy‑efficient than the prior generation.

Lambda Enhancements

New EC2 Instance Types

• Memory‑optimised instances powered by 5th‑Gen AMD EPYC processors.
• Up to 5 GHz clock speeds and 3 TiB of RAM—targeted at heavy databases, EDA tools, and similar workloads.

AI Services & Models

Amazon Nova 2 Family

All Nova 2 models support adjustable intensity levels (dial up reasoning depth when needed, keep it light for simple queries).

Amazon Nova Forge

• “Build your own frontier model.”
• Start from Nova checkpoints, blend proprietary data with Nova’s training data, and obtain a custom model that combines Nova’s capabilities with your domain knowledge.
• Reddit reportedly built its own model using Forge.

Nova Act

• Browser‑automation agents powered by a custom Nova 2 Lite model.
• Claims ≈ 90 % reliability on browser‑based tasks—useful for web scrapers and automation tools.

Amazon Bedrock Updates

• > 100 foundation models, including 18 new open‑weight models added in December.
• Reinforcement fine‑tuning: feedback‑driven training that delivers ≈ 66 % accuracy gains without massive labelled datasets.

Amazon Bedrock AgentCore

• SDK downloads: > 2 M in the first 5 months.

Core Components

• Runtime – Session isolation, bidirectional streaming for voice agents.
• Memory – Includes episodic memory for agents that learn from experience.
• Gateway – Convert existing APIs to MCP‑compatible tools.
• Identity – OAuth integration and secure token storage.
• Observability – CloudWatch dashboards for agent monitoring.
• Policy – Real‑time tool‑call interception using Cedar policies.
• Evaluations – 13 built‑in evaluators for automated testing.

Bottom Line

2025 was a transformative year for AWS:

• AI agents are now first‑class citizens on the platform.
• Custom silicon (Graviton 5, Trainium 3) pushes performance and efficiency to new heights.
• Multicloud capabilities are being baked into the core offering.

If you’re a developer or architect, the most practical takeaways are the Lambda Durable Functions, Managed Instances, and the Nova 2 model family—these will likely impact day‑to‑day workloads the most. Keep an eye on AgentCore as it matures; it could become the de‑facto framework for building production‑grade AI agents on AWS.

Amazon Q Developer

• Languages: C#, C++, and 11 additional languages for customisation.
• GitLab Duo integration: GA.
• GitHub integration: preview – no AWS account required.
• MCP support in the CLI.
• Conversation history: persists between sessions.
• Pro Tier: available in Frankfurt for EU data residency.

Kiro IDE

AWS released Kiro, an agentic AI IDE that Amazon now uses internally. Built on VS Code, it features spec‑driven development where you write requirements in markdown and the agent scaffolds everything.

• One internal project reportedly went from 30 developers over 18 months to 6 developers over 76 days.
• Early signs are promising, though I’m still evaluating it.

Database Savings Plans

A single, flexible commitment that applies across RDS, Aurora, DynamoDB, ElastiCache, Neptune, and DocumentDB. No more juggling separate Reserved Instance portfolios per engine.

• Up to 35 % cost reduction with a one‑year commitment.

Aurora DSQL

• Cluster creation now takes seconds instead of minutes – great for rapid prototyping and testing.

RDS Storage Expansion

• SQL Server and Oracle now support up to 256 TiB of storage (up from 64 TiB).
• 4× improvement in IOPS and I/O bandwidth.
• Makes migration from large on‑prem databases easier.

OpenSearch Enhancements

• GPU‑accelerated vector indexing – 10× faster at one‑quarter the cost.
• Auto‑optimised vector indexes automatically evaluate different KNN algorithms to balance recall quality against query performance.

AWS Interconnect

A partnership with Google Cloud to offer managed, high‑speed private connections between the two platforms.

• Provision dedicated bandwidth on demand and establish connectivity in minutes.
• Includes quad‑redundancy and MACsec encryption.
• Microsoft Azure support is slated for 2026.

“Multicloud is mostly marketing, but this is actually useful.” – my take

Route 53 Global Resolver (Preview)

Secure anycast DNS resolution that simplifies hybrid DNS management – one service instead of managing resolvers in each VPC.

Amazon S3 Vectors (GA)

• Native support for storing and querying vector embeddings.
• Scales up to 2 billion vectors per index (40× the preview capacity).
• Supports up to 20 trillion vectors per bucket.
• Reduces costs by up to 90 % compared with specialised vector databases.

Great for RAG, semantic search, and agentic workloads – eliminates the need for a separate vector DB (e.g., Pinecone).

S3 Tables

• Built‑in Intelligent‑Tiering support and replication for Apache Iceberg‑native tables.
• Simplifies analytics on S3 without complex ETL pipelines.

FSx for NetApp ONTAP

• Now integrates with S3, allowing file‑system data to be accessed via S3 APIs.
• Useful for plugging existing storage into analytics and ML services without copying data around.

Amazon EKS Capabilities

Fully managed platform capabilities for workload orchestration and cloud resource management.

• Aims to eliminate infrastructure maintenance while maintaining enterprise‑grade reliability.

If you’re running vanilla Kubernetes and spending too much time on cluster management, this might help.

ECS Express Mode

Simplified container deployments for ECS.
I haven’t tried this one yet, but it’s on my list.

Product Lifecycle Page (2025)

AWS introduced a consolidated page for all service availability information – about time.

Key Deprecations to Be Aware Of

If you’re using any of these, now’s the time to plan your migration.

The CodeCommit Reversal

In a rare move, AWS reversed the CodeCommit deprecation in November 2025 after listening to customer feedback.

• CodeCommit is back to full General Availability with new features planned:
  • Git Large File Storage (early 2026)
  • Regional expansion to additional regions starting Q3 2026

The reversal acknowledged CodeCommit’s deep IAM integration, VPC endpoint support, and seamless connectivity with CodePipeline—especially valuable for regulated industries. AWS explicitly apologized for the inconvenience caused to customers who had begun migration planning.

2025 – Patterns & Takeaways

1. AI agents are the new compute primitive

   • From servers → containers → functions → agents.
   • AWS is betting big that autonomous systems will act on our behalf.

2. Custom silicon matters

   • Continued investment in Graviton and Trainium to out‑price and out‑perform commodity hardware for specific workloads.

3. Multicloud is becoming practical

   • The Google partnership shows AWS recognises legitimate multicloud needs. Expect more interoperability announcements in 2026.

4. Developer experience is a priority

   • Q Developer, Kiro, and various IDE integrations demonstrate AWS’s commitment to AI‑assisted development tools, not just infrastructure.

Whether you’re excited or exhausted by the pace of change, 2025 was undeniably a significant year for AWS. And if the roadmap announcements are any indication, 2026 will … (the original text cuts off here).

---

What announcements mattered most to you? I’d love to hear what you’re planning to try first.