AWS in 2026: S3 Vectors, Nova Act, and the AI Infrastructure Wave

AWS’s 2026 story is really one story: AI workloads going mainstream, and the infrastructure reshaping around them. Vector search moves into object storage, agents become first-class citizens, and inference gets its own routing layer. Here are the releases that actually matter — and what to do about them.

S3 Vectors: vector search at object-storage economics

Amazon S3 Vectors reached general availability across 14 regions: read latency around 100ms or less on frequent queries, up to 100 results per query, and 1,000 PUTs per second per index. The strategic point matters more than the specs — vectors move from specialized databases to S3’s scale and cost model. If you were holding off on semantic search because of infrastructure cost or operational overhead, re-run the numbers: the economics just changed.

Nova Act: agents that use the browser

Amazon Nova Act is now generally available — agents that automate UI workflows directly in the browser. Paired with Nova Forge, which lets you build custom frontier models from Nova checkpoints with your own data, AWS now covers the agent stack from model customization to action. The “agent” is no longer a demo concept; it is a deployable workload with managed infrastructure underneath it.

Inference gets a routing layer

The SageMaker HyperPod Inference Gateway adds intelligent routing for large language models, and Bedrock AgentCore memory brings long-term memory ingestion to agents. Inference is no longer just “call the model” — it is becoming a tier with routing, memory, and policy, the same way compute grew load balancers and caching layers. Architect accordingly.

The platform keeps moving

Beyond the AI headlines, the platform fundamentals advanced:

• AWS Lambda managed runtimes entered public preview — less runtime maintenance for serverless teams.
• RDS for PostgreSQL gained post-quantum TLS support — start planning your cryptographic agility now, not when it becomes urgent.
• New EC2 families: T8i for burstable workloads and R9g/R9gd for memory-optimized Graviton workloads.

What to do about it

Two practical takeaways. First, if semantic search is on your roadmap, S3 Vectors deserves a proof of concept — the cost model shifted under the old assumptions. Second, treat agentic workloads as first-class architecture now: memory, routing, and tool use are infrastructure concerns, not app-layer hacks. The teams that design for agents deliberately will spend far less retrofitting later.

The cost lens

Every AWS announcement is also a pricing announcement in disguise. S3 Vectors matters because vector workloads previously needed a specialized database — provisioned, scaled, and paid for separately. Moving vectors into S3 collapses a line item into storage economics you already understand. Similarly, Lambda managed runtimes in preview promise less operational toil, and Graviton-based R9g/R9gd instances continue the pattern of better price-performance for memory-heavy workloads. When you evaluate the 2026 releases, model total cost of ownership, not just features: the cheapest vector database is increasingly the object store you already pay for.

Post-quantum readiness starts now

RDS for PostgreSQL gaining post-quantum TLS support deserves more attention than it will get. Quantum-safe cryptography is a migration measured in years, not sprints: certificates, client libraries, and compliance frameworks all have to move. AWS putting it in a managed database means you can start experimenting with hybrid handshakes on real workloads today. Add it to your security roadmap as a track, not a ticket — the teams that start early will not be the ones scrambling later.

For teams building on Bedrock or SageMaker, the practical shift is treating inference like any other distributed system: route requests intelligently, keep memory close to the workload, and set policy at the gateway. AgentCore memory means your agents can now remember across sessions without you building a memory store from scratch — one fewer bespoke system to operate.

“In 2026, the question is not whether to run AI workloads on AWS, but which managed layer absorbs each piece.”

— DesignsToDeploy engineering notes

We architect cloud systems for clients on AWS every week — from vector search proofs of concept to production agent infrastructure. If you are scoping semantic search or agentic features, we can map the right services to your budget at designstodeploy.dev.