What We Saw At AWS re:Invent 2025

The premier cloud event of the year, AWS re:Invent, was a week-long cloud customer confab. The central message: AI is just another workload, best consumed as a managed service on AWS’s public cloud platform (or your own data center with AWS’s AI Factories). The goal? Position AWS as a newly transformed, AI-native cloud that’s ready right now. Keynote speakers touted the big Project Rainier AI data center (already operational), a raft of new AI-enabled services now generally available, and many more in preview. AWS’s “everyday AI” positioning prominently featured AgentCore — with a promise to allow customers to create agentic “teammates” — along with enabling technologies centered on the Bedrock AI platform, which got a slew of third-party models and the second generation of AWS’s in-house Nova models. AI development got airtime with a spruced-up version of the Kiro integrated developer environment. While the theme was business as usual, AWS CEO Matt Garman took several opportunities to state just how big AWS’s business is: an estimated $132 billion in 2025 with 20% annual growth and a projected goal of $300–400 billion in annual revenue. What wasn’t featured as prominently? AWS didn’t play up its $38 billion deal to run OpenAI workloads, instead promoting Nova models for enterprise use cases. Similarly, key partners such as Anthropic (which received a $4 billion investment from AWS) and NVIDIA took a back seat in AWS messaging. NVIDIA was advertized as just part of a broader portfolio of the AI infrastructure story, alongside its own Trainium custom silicon. Some announcements that would have been headline-making not long ago, such as the AWS Interconnect offering for multicloud integration, got less attention relative to AI. Less pleasant topics — such as the big October outage triggered by DynamoDB — went undiscussed. Here’s a closer look at key announcements and developments:

• AWS is a grocery store for AI models, and Nova is the store brand. Bedrock now includes 18 fully managed open weight models from providers ranging from Google and Meta to Alibaba and Mistral, with Kimi and MiniMax for regional relevance in China. Amazon’s Nova models bring multimodality and audio capabilities for business functions. Nova Forge allows customers to modify models by adding their own data for pretraining. Those offerings, plus Bedrock’s new reinforcement learning feature, lower the barrier for developers and data scientists to align models to the business.
• Black box “vibe-coding” is out, while “spec-driven” code assistance is in. The “chat and hope” model of AI-assisted development is giving way to structured, opinionated agentic development workflow. Thus, Kiro transforms natural language into specification artifacts before a single line of code is written by its “frontier agents.” This philosophy extends to AWS Transform custom, which pushes this spec-driven approach to allow enterprises to set code migration rules before unleashing agents to burn down technical debt.
• AI Factories bring AWS into customers’ data centers. AWS’s grudging acceptance of data center workloads via its Outposts offering has given way to an enthusiastic push for AWS AI Factories, addressing sovereignty concerns while AI adoption accelerates. It’s likely that well-resourced enterprises in regulated sectors will adopt the offering, as enterprises in healthcare and financial services refuse to build out cloud-resident AI if it means loss of autonomy and control. While AWS manages the hardware, customers still need robust facilities (power, cooling, physical security). Smaller organizations may struggle to take advantage without colocation support.
• Neuro-symbolic guardrails will shape AWS AI offerings. AWS AgentCore Policy and specialized frontier agents for development (Kiro), security, and DevOps will all be safeguarded by neuro-symbolic guardrails via Cedar, AWS policy language that links the neural networks of large language models and symbolic language. Policies are critical for adding (if not yet ensuring) deterministic oversight in nondeterministic AI environments. Even so, AgentCore policies need broader integration to deliver enterprise-grade governance. Frontier agents help organizations accelerate but require greater ease of use for context engineering and native integration with business processes. These are essential building blocks, but success will require additional policy design and significant data and process reengineering.
• Sovereignty strategy takes a different approach to jurisdictional independence. The upcoming launch of the AWS European Sovereign Cloud (ECS) adds to other sovereignty initiatives such as the UAE sovereign launchpad and the AWS dedicated local zones. The ECS will be globally accessible and will solve the problem of advanced data residency and controls. AWS’s approach to sovereignty differs from its main competitors, Google Cloud and Microsoft, which use partnerships and joint ventures with local companies to address sovereignty requirements and the perils of a “kill switch.”
• AWS entices US government agencies with new infrastructure build-outs. US government agencies will see the first-ever genAI and high-performance computing purpose-built infrastructure, in the form of 1.3 gigawatts in compute capacity from a promised 10-year, $50 billion investment. This compute capacity will be allocated across four US regions — Commercial, Top Secret, Secret, and GovCloud — and will be available in 2026 for both Commercial and GovCloud customers. This announcement follows on the heels of AWS OneGov agreement with the General Services Administration to provide $1 billion credits to federal agencies, highlighting AWS’s moves to maintain access to the potential $80 billion government market.
• Sustainability gets new attention. AWS now tracks scope 3 emissions in its carbon footprint tool, showing customers why emissions fluctuate and how different AWS services impact emissions. On top of region-based emissions, AWS now allows location-based emissions, overcoming the challenge of reporting net-zero emissions only as an effect of renewable energy purchases. Once the reporting methodology gets more established, the plan is to move from reactive reporting to proactive optimization, addressing the problem of overprovisioning.
• Amazon Connect keeps growing. At its most recent earnings call, Amazon reported that the Connect contact-center-as-a-service solution had surpassed $1 billion in annual revenue. At re:Invent, it announced a number of new capabilities for the product. Highlights include ever more human-sounding customer conversations enabled by a new integration with AWS’s Nova Sonic speech modules, as well as support for third-party automatic speech recognition and text-to-speech systems such as Deepgram and ElevenLabs. The announcement included 29 new features, notably agentic AI for better customer self-service and new conversational interfaces for system administrators.
• Commodity cloud gets more clout. Good old network, compute, and storage doesn’t generate many headlines these days. Nevertheless, the legions of cloud engineers at re:Invent came home with a virtual armful of new services to consider, ranging from implementation headache relief (EKS Capabilities) to added compute muscle (EC2 C8a and X8aedz instances based on fifth-gen AMD EPYC processors for compute and memory-heavy workloads). Serverless got new juice, too, with Lambda Managed Instances to streamline EC2 and Lambda durable functions for complex and long-running applications for use cases such as agentic deployment.
• Data and storage capabilities get a refresh for AI readiness and scale. The general availability of S3 vectors aligns with the convergence of data, data management, and storage into data infrastructure for AI. Amazon OpenSearch is now retooled for better vector databases and GPU acceleration. For enterprise-friendly object storage, Amazon FSx for NetApp ONTAP now integrates with S3 for more AI-readiness efforts. AWS is pushing toward a future where lakehouses, databases, and vector retrieval operate as a seamless fabric optimized for AI workloads. New storage efficiencies, larger object limits, and faster data operations position S3 as the backbone for retrieval-augmented generation and real-time agentic use cases.
• Database users get a price break. While re:Invent 2025 didn’t bring major new database features, it delivered meaningful cost savings. The new Database Savings Plans offer up to 35% reductions across seven services (Aurora, RDS, DynamoDB, DocumentDB, Neptune, ElastiCache, and Timestream) with flexible, cross-service commitments. AWS also announced up to 55% lower costs for Amazon RDS for SQL Server using optimized CPU configurations on the new M7i and R7i instance families, dramatically cutting vCPU-based licensing overhead. These updates give enterprises more predictable spend, greater architectural flexibility, and substantial cost efficiency, making it easier to scale data workloads for AI.
• Security Hub steps up. AWS announced a more advanced Security Hub, which aggregates signals from Amazon GuardDuty, Amazon Inspector, AWS Security Hub Cloud Security Posture Management, and Amazon Macie. It still falls short, however, of effectively supporting non-AWS clouds such as Microsoft Azure or Google.
• AWS shows observably agentic ambitions. AWS embeds monitoring, evaluation, and policy enforcement into AgentCore, positioning observability as a trust fabric rather than a set of dashboards. Integration with OpenTelemetry and CloudWatch introduces continuous evaluations, episodic memory, and anomaly detection — key differentiators for governance and performance assurance. This aligns with Forrester’s vision of AIOps platforms, with observability becoming a dynamic control layer for autonomous operations. AWS now provides the foundational blocks for AIOps maturity, but success will depend on how quickly organizations operationalize and scale these capabilities.