The News
At an AWS re:Invent 2025 event preview, AWS laid out an AI-first roadmap centered on agentic AI, the Nova 2 model family, a new Forge service for custom frontier models, and deeper full-stack integration from NVIDIA-powered clusters and Trainium chips up through Bedrock and new agent frameworks like AgentCore and “frontier agents.” The briefing framed this re:Invent as a pivot from generic GenAI experimentation to production-scale agentic systems that span clouds, sovereign environments, and enterprise workloads.
Analysis
Agentic AI Becomes the Center of Gravity
AWS opened by emphasizing record attendance (more startups, more developers, more senior executives than in prior years) and tied that directly to an inflection point in agentic AI. The message was that re:Invent 2025 is about moving from “AI curiosity” to AI-first reinvention, with agents as the organizing principle.
The preview made it clear that AWS sees the next decade of cloud growth emerging from billions of agents (small and large, internal and external) rather than simply more microservices. For developers, that means agent patterns, data governance, and safety will increasingly dictate how systems are designed and deployed.
Infra, Sovereignty, and Model Choice
AWS spent significant time reinforcing its “full stack” story, from silicon to large clusters to managed model platforms. The company highlighted deep work with NVIDIA, the new AWS AI Cluster concept for sovereign and public-sector deployments, and new Trainium 3 and Trainium 4 generations aimed at frontier-scale training.
For builders, the important part is that this stack is being framed as AI-first, not just cloud-first. The AWS AI Cluster is meant to bring AWS’s operational practices into customer-controlled facilities where data residency and sovereignty are non-negotiable. That acknowledges a reality many teams face: some of the most sensitive AI workloads cannot live solely in shared public regions.
On the model side, AWS is growing Bedrock into a broad multi-model hub, with 18+ new open-weight models plus the Nova 2 family. The promise is that teams can mix and match models like Nova, third-party, and open-weight without redesigning their underlying infrastructure each time. In a market where 54.4% of organizations already run hybrid environments and large enterprises are standardizing on multi-model strategies, that flexibility matters.
Custom Frontier Models as a Service
The Forge was one of the more strategically interesting previews. It is pitched as an “open training model capability” that sits between fine-tuning and building a model from scratch. Customers start from Nova training checkpoints, blend their own proprietary data with curated data, and produce a custom frontier model that still runs within Bedrock.
AWS described the flow in three main steps:
- Access Nova checkpoints (early, mid, late) as a starting point.
- Blend proprietary data with curated training data to shape the model.
- Deploy the resulting model directly into Bedrock with its existing security and governance.
This approach is aimed at a real gap. Many enterprises want models that deeply understand their domain, but don’t want to lose the generalized capabilities of a frontier model or take on the cost and complexity of full-from-scratch training. ECI data shows 86% of organizations view unifying and activating proprietary data as critical to AI success, and a significant subset worry about over-fine-tuning degrading model behavior.
For developers, Forge could become a way to justify investing in “house-branded” models for high-value agentic workloads, particularly in regulated industries, without rebuilding the entire stack or rolling their own MLOps pipelines.
AgentCore, Policy, Evaluations, and Frontier Agents
The preview also clarified how AWS wants teams to build and operate agents. AgentCore is positioned as the baseline runtime and SDK, already seeing strong early adoption. On top of that, AWS is introducing policy and continuous evaluations to make agents safer and more predictable at scale.
The agent stack was summarized as:
- AgentCore as the foundational framework for building agents.
- Policy at the gateway, enforcing constraints on what agents can do and which tools or actions they can invoke.
- Continuous evaluations, with pre-built evaluators and ongoing monitoring to detect drift or quality issues.
- Frontier agents, a category of fully autonomous, scalable, long-running agents for development (Kiro), security, and DevOps use cases.
This is directly aligned with what we’re seeing in the market. Organizations are leaning heavily into automation. 71% already use some form of AIOps, and 45.4% want observability that can detect misconfiguration, access violations, and drift in near real time.
For developers, the message is that building a serious agent system is not just “calling an LLM.” It involves identity, state, policies, evals, and long-running workflows. AWS wants those concerns to be handled by platform components rather than bespoke glue code that has to be re-implemented team by team.
Data, Storage, and Economics for AI-Scale Apps
The preview closed with a set of data, storage, and security updates that quietly matter a lot to AI-native architectures. Increasing S3’s max object size to 50 TB opens the door to ultra-large multimodal artifacts, dense logs, and training datasets that don’t need awkward sharding strategies. New database features and savings plans are pitched as cost-control levers for AI-heavy workloads that are naturally data-hungry.
Security enhancements such as extending GuardDuty coverage and adding new Security Hub capabilities fit into the same story. As organizations scale agentic workloads, detection and response need to evolve alongside them. With 60.7% of organizations planning to increase cloud infrastructure spend and 62.7% prioritizing security and compliance, the economic and risk profiles of AI-native applications are becoming executive-level concerns, not just technical implementation details.
For developers, this means storage layouts, indexing strategies, and observability pipelines must be designed assuming multi-terabyte artifacts, cross-region data flows, and agents that operate over long time horizons.
Looking Ahead
The event preview laid out AWS’s view of the AI-native cloud as a tightly connected stack with accelerated hardware, sovereign-capable clusters, a broad model catalog, custom frontier training via Forge, and an opinionated agent platform with policy and evals built in. It’s a vision where billions of agents sit on top of a governed, multi-model, multi-cloud fabric.
In the near term, the key question for the industry is how quickly these capabilities translate into repeatable patterns developers can adopt without heroic effort. If Nova 2, Forge, AgentCore, and frontier agents integrate cleanly into real-world pipelines, AWS could become the default place to design, govern, and run agentic systems end to end. If not, teams will continue to assemble their own stacks across multiple vendors and clouds, using AWS as one substrate among several.
Either way, the next phase of cloud isn’t just “more compute and storage,” it’s about reliable, governed, and explainable agents operating at scale, and developers will be the ones stitching those capabilities into real applications.
Agentic AI Becomes a Full Stack, Not a Feature
Akuity Expands GitOps Delivery to Multi-Cloud & Multi-Environment Targets
LocalStack Hits Record Adoption as Local-First Cloud Development Accelerates
Inside Amazon’s New Framework for Measuring Developer Productivity
Agentic AI, Nova Forge & Cloud-Scale AI
