AWS re:Invent 2025 Key Announcements: What They Mean for the Future of Cloud, AI & Enterprise Tech

AWS re:Invent 2025 made one thing very clear: AWS is fully committing to an AI-first, agent-driven, and hybrid-ready future.

With major announcements across generative AI, infrastructure, privacy, cost optimisation, and hybrid cloud, the keynote set the direction for how modern applications will be built and scaled over the next decade.

Below is a summary of the key announcements and what this means for developers, enterprises, and the cloud industry as a whole.

#	Announcement	Area	Why it matters
1	P6e GB300 NVIDIA GPU instances	Compute / AI	New EC2 instances using NVIDIA GB200 NVL72 systems, offering ~20× the compute of prior P5en for huge training and inference jobs - aimed at frontier-scale AI and agents.
2	AWS AI Factories	Hybrid AI infra	Lets customers deploy AWS AI infrastructure (Ultra servers, Trainium, Bedrock) directly into their own data centres, giving "private AWS-like regions" for regulated or on-prem workloads.
3	Trainium 3 Ultra servers GA	AI accelerators	Third-gen Trainium Ultra servers become generally available with big boosts in compute, memory bandwidth and energy efficiency, turning racks into AI supercomputers for training LLMs.
4	Trainium 4 preview	AI accelerators	Next-gen Trainium is announced with large jumps in compute and bandwidth over Trn3, targeting future "absurdly large" frontier models and long-term AI roadmap planning.
5	Mistral Large & Mistral 3 open-weights in Bedrock	Models / Inference	High-performance open-weights models from Mistral are added to Amazon Bedrock, giving customers more choice and flexibility for both heavy reasoning and efficient edge/latency-sensitive use cases.
6	Amazon Nova 2 model family (Lite, Pro, Sonic)	Models / Inference	New foundation models optimised for cost and quality: Lite for fast, cheap tasks; Pro for complex reasoning and agent workflows; Sonic for low-latency speech-to-speech conversations - AWS's main answer to other frontier models.
7	Amazon Nova 2 Omni	Multimodal AI	A unified multimodal model that ingests text, images, audio, and video and can output both text and images, simplifying scenarios like watching a video presentation and generating summaries plus visuals in one shot.
8	Amazon Nova Forge ("Novella" training)	Custom model training	A training service that lets customers start from Nova checkpoints and train "open training models" with their own data (mid-training, not just fine-tuning), producing custom "Novella" models tailored to their domain.
9	Policy in AgentCore	Agentic AI / Governance	A policy engine for Amazon Agent Core that defines what agents are allowed to do, with which tools and under what conditions - similar to IAM for agents - giving deterministic safety controls beyond prompts.
10	AgentCore Evaluations	Agentic AI / QA	Built-in evaluation tools for Agent Core with pre-made and custom metrics (correctness, safety, usefulness) to continuously score and monitor agent behaviour in production - essentially QA for agents.
11	Kiro Autonomous Agent	Dev tooling / Agents	An autonomous development agent (built on Kiro) that can take a goal, plan work, update code across repos, write tests, and open PRs with minimal supervision - targeted at long-running engineering tasks like big refactors.
12	AWS Security Agent	Security / Agents	A specialised security agent on AgentCore that reviews code and configs, flags policy/security violations and suggests fixes, integrating with pipelines to act as a continuous compliance assistant.
13	AWS DevOps Agent	Ops / Agents	An AgentCore-based DevOps agent that automates provisioning, CI/CD changes, config checks and rollbacks, effectively serving as an always-on extra DevOps team member.
14	X8i memory-optimised instances (Intel Xeon 6)	Compute	New X-family EC2 instances with custom Intel Xeon 6 chips, delivering up to ~50% more memory for big in-memory workloads like SAP HANA or large databases.
15	Next-gen AMD EPYC memory instances (3TB RAM)	Compute	New AMD-based instances with up to 3 TB RAM, giving another option for very large memory-bound applications at competitive price-performance.
16	C8a instances (AMD EPYC)	Compute	CPU-optimised instances using latest AMD EPYC, promising around 30% better performance for compute-heavy tasks such as game servers or batch processing.
17	C8iNE instances (Intel + Nitro v6)	Compute / Networking	Network-enhanced compute instances combining Intel Xeon 6 with Nitro v6 to deliver about 2.5× better packet performance per vCPU, aimed at security appliances, firewalls and network-intensive services.
18	M8 AZN high-clock instances	Compute / Low-latency	New M-family instances with very high single-threaded clock speeds, aimed at latency-critical workloads like gaming, real-time analytics and trading systems.
19	EC2 M3 Ultra Mac	Apple / Dev	One of two new Mac instances, providing Apple Silicon-based environments for building and testing macOS/iOS apps in the cloud with more power and scale.
20	EC2 M4 Max Mac	Apple / Dev	The second new Mac instance type using the latest Apple chips, giving even higher performance for iOS/macOS CI pipelines and multi-platform app shops.
21	Lambda Durable Functions	Serverless	Lambda is extended to support long-running, stateful functions that can run for hours or days with resumability and retries - ideal for workflows waiting on agents, human approvals or long processes.
22	S3 max object size increased to 50 TB	Storage	S3's individual object limit jumps from 5 TB to 50 TB, simplifying storage of huge datasets, high-res media, and very large model checkpoints without chunking.
23	S3 Batch Operations 10× faster	Storage / Data ops	S3 Batch Operations are significantly sped up, reducing time and cost for bulk tasks like tagging, copying, and transforming data at petabyte scale.
24	Intelligent tiering for S3 Tables (Iceberg)	Storage / Analytics	S3 Tables (Apache Iceberg) gain intelligent tiering, automatically moving colder table data to cheaper storage classes and potentially cutting costs by up to ~80%.
25	S3 Table replication across regions/accounts	Storage / DR	S3 Tables can now be replicated across regions and accounts, enabling globally consistent query performance and simpler multi-region analytics setups.
26	S3 Access Points for FSx for NetApp ONTAP	Storage / Integration	S3 Access Points are extended to FSx for NetApp ONTAP so ONTAP file data can be accessed like native S3 objects, easing hybrid file/object workflows.
27	S3 Vectors GA	Storage / Vector DB	Native vector storage in S3 (S3 Vectors) becomes generally available, designed to hold and search trillions of embeddings with much lower cost than many bespoke vector databases.
28	GPU-accelerated vector indexing for OpenSearch	Search / AI	OpenSearch adds GPU acceleration for building vector indices, shrinking indexing time by around 10× and cost by ~75%, which is important for large-scale semantic search.
29	EMR Serverless - no local storage config needed	Analytics / Big data	EMR Serverless clusters no longer require you to provision local storage, removing a major configuration hassle and making EMR closer to "pure" serverless big-data processing.
30	GuardDuty support for ECS	Security / Threat detection	Amazon GuardDuty's threat detection expands to ECS workloads, enabling managed anomaly and malware detection for containerised apps.
31	Security Hub GA with new analytics	Security / Posture mgmt	AWS Security Hub becomes generally available with real-time risk analytics, trend views and cleaner pricing, centralising security findings across services.
32	Unified CloudWatch log store	Observability	CloudWatch introduces a unified data store that aggregates logs from AWS services and third-party tools (like Okta, CrowdStrike) into one searchable, analytics-ready location.
33	RDS storage expansion for SQL Server & Oracle	Databases	Amazon RDS lifts storage limits up to 256 TB for SQL Server and Oracle, increasing capacity and I/O throughput for very large enterprise databases.
34	Configurable vCPU counts for RDS SQL Server	Databases / Licensing	You can now set custom vCPU configurations for RDS SQL Server, helping tune instance sizing to optimise Microsoft licence spending.
35	RDS support for SQL Server Developer Edition	Databases / Dev & test	RDS adds support for SQL Server Developer Edition at zero licence cost, making it easier and cheaper to build and test SQL Server-backed apps in the cloud.
36	Database Savings Plans	Databases / Pricing	New Savings Plans for databases offer up to about 35% discounts across multiple engines, finally giving a unified, predictable cost model for long-running DB workloads.