AWS re:Invent dropped a massive wave of AI updates on Day 2. Matt Garman, CEO, AWS held zero punches back. He delivered a roadmap for agentic AI at unstoppable enterprise scale with new Nova models, durable functions, massive scale, analytics evolution, S3 reinvention, and agentic architectures.
Every Day 2 talk hammered home systems that reason like superbrains, scale to the stars, and deliver massive ROI.
Durable workflows conquering impossible marathons. Proven scale obliterating records. Analytics turbocharged for AI dominance. Storage simplicity that's pure genius. Autonomous agents unlocking big wins wins.
The announcements and launches add an impressive host of capabilities to the powerful enterprise AWS AI ecosystem.
Here’s the quick rundown of every big announcement from a very big day.
The Big AWS re:Invent Day 2 Launches
Nova 2 Model Family
Amazon Nova 2 introduces a suite of frontier foundation models optimized for cost, speed, and versatility, powering diverse AI experiences from reasoning to multimodal tasks.
- Nova 2 Lite: Fast, cost-effective reasoning model excelling in instruction following, tool calling, code generation, and document extraction outperforms Claude Haiku 4.5, GPT-5 Mini, and Gemini Flash 2.5 at leading price-performance.
- Nova 2 Pro: Advanced reasoning for complex workloads, topping benchmarks in agentic tool use vs. GPT-5.1, Gemini 3 Pro, and Claude 4.5 Sonic.
- Nova 2 Sonic: Next-gen speech-to-speech model enabling real-time, human-like multilingual conversations with low latency and expanded language support ideal for telephony and interactive apps.
- Nova2 Omni: Industry-first multimodal model handling text, image, video, and audio inputs with text/image outputs unifies reasoning across modalities for tasks like keynote summarization with visuals.
Our experts have started building a comprehensive guide to Nova 2 and a deep dive guide to Nova 2 Omni. We'll update them live as we begin building more solutions with these frontier models.
Nova Forge
New service providing exclusive access to Nova training checkpoints, allowing customers to blend proprietary data with Amazon-curated datasets during pre-training creating custom "novellas" that embed domain knowledge without losing core reasoning capabilities. Supports remote reward functions and reinforcement fine-tuning for production-ready frontier models; upload directly to Bedrock.
Lambda Durable Functions
Enhancement to AWS Lambda enabling durable, long-running workflows (up to one year) with managed state, built-in error handling, and automatic recovery. Program "steps" for logic/retries and "waits" for pauses (e.g., AI agents, human approvals) pay only for active compute, scale-to-zero during idle; supports Python/Node.js via SDK and deployment via SAM/CDK.
AWS AI Factories
Dedicated, AWS-managed AI infrastructure deployed in customer data centers like private AWS regions for exclusive use. Includes Trainium/NVIDIA GPUs, SageMaker, Bedrock; meets sovereignty/compliance needs while leveraging AWS reliability (e.g., Saudi Arabia's Humane AI zone).
Trainium3 Ultra Servers (GA) & Trainium4 Preview
- Trainium3: 3nm chips in Ultra Servers with 144 chips per rack (362 FP8 petaflops, 700+ TB/s bandwidth); 4.4x compute, 3.9x memory bandwidth, 5x tokens/megawatt vs. Prior 5x efficiency on GPT-OSS 120B. Over 1M chips deployed, multi-billion business.
- Trainium4: Upcoming with 6x FP4 compute and 4x memory bandwidth over Trainium3 for the largest models.
AgentCore Updates (Bedrock)
Amazon Bedrock AgentCore platform accelerates production agents with composable services (Runtime, Gateway, Policy preview, Memory, Identity, Evaluations preview, Observability, Code Interpreter, and Browser). Features real-time policy enforcement (natural language to Cedar), session isolation (up to 8 hours), tool discovery, and CloudWatch metrics for quality (correctness, safety). Framework-agnostic, secure scaling used by Ericsson, Thomson Reuters.
Amazon QUIC
Unified enterprise AI productivity assistant integrating data, BI, research, and workflow automation streamlines tasks across tools for faster decision-making (likely the Amazon Q extension; previewed for business users).
Amazon Bedrock Model Expansion
Doubled models to include 18+ new ones: Mistral Large (5x parameters, 2x context), Mistral 3 (edge/single-GPU), Google's Gemma, and NVIDIA Nemotron plus proprietary/open-weights for broad choice. 50+ customers processed 1T+ tokens each; Trainium powers majority inference.
Amazon Development Agents
Kiro development agents, AWS Security Agent, and AWS DevOps Agent are now available in Preview to accelerate development.
Upto 100 Kiro seats free for startups for one year. Apply within the next one month.
Titans take the stage: how leading companies lean on AWS
Garman welcomed industry leaders who showcased how real-world AI agents are already delivering meaningful results.
John Kodera, Chief Digital Officer at Sony, highlighted how AWS has supported the company’s evolution from the early PlayStation Network era to today’s global digital experiences. He spoke about AWS’s ability to scale reliably, power data-driven services, and help Sony deepen connections between creators and fans around the world.
Shantanu Narayen, Chair and CEO of Adobe, emphasized how AWS underpins the company’s continued shift toward AI-enabled creativity and productivity. From pioneering cloud subscriptions to delivering secure, scalable access to advanced models through Adobe’s products, he noted that AWS continues to match Adobe’s pace as AI demand grows across their customer base.
May Habib, CEO and Co-founder of Writer, demonstrated how the company uses Palmyra X5 agents on Amazon Bedrock to support enterprise customers such as Uber, Accenture, and Vanguard. She highlighted Writer’s focus on strong governance and data security, explaining that hundreds of enterprises rely on AWS’s isolation capabilities and model offerings to build grounded, production-grade AI agents.
Their stories are not very different from GoML customers like Atria, Olympian Motors, Lyzr, Miden and others who have scaled their AI workloads and customer experiences on AWS stack.
Amazon's AI scale
Amazon flexed Prime Day god-mode: 200M+ users, 9B packages. Rufus AI blasted 3M tokens/min on 87K Trainium chips <1ms latency, 4x cost carnage, 60% purchase surges!
21K+ Bedrock agents slashed $2B costs; one nuked delivery defects 74% with S3/DynamoDB. Zoox robotaxis devour petabytes on SageMaker HyperPod (10K+ GPUs) for Vegas autonomy. Graviton/DynamoDB/CloudFront crushed 3T+ requests via continuous batching
The message: AWS primitives free builders to innovate while handling scale.
Harnessing analytics for humans and AI
Mai-Lan Tomsen Bukovec outlined analytics shifts for agentic apps. Open Table Formats like Iceberg enable lakehouses; services converge for unified governance.
Spark 3.5.6 boosts Iceberg across EMR/Glue/Athena with AI migration agents. Kinesis streaming powers real-time AI; SageMaker Catalog adds vector search. FINRA and others future-proof data oceans for human-AI decisions.
Frictionless data foundations accelerate agentic infrastructure.
S3 keeps evolving
S3 evolved with 1,000+ updates since 2020, prioritizing reliability through conservative iteration. Conditional PUT prevents data loss in distributed writes, handling millions per minute.
Object sizes hit 50 TB; Conditional Copy/Delete simplify apps. These enable lock-free architectures for logs and elections.
S3's philosophy: Scale silently so customers focus on value.
AI agents in action
Nandi compared agentic AI to 2013 cloud skepticism, now proven at Amazon with $2B savings and 4x developer gains. Agents automate complex tasks for exponential value.
ASAP hit 91% first-call resolution, 77% cost cuts; NoHarm boosted hospital capacity 8x, saving $30M+. AgentCore delivers policy, memory, and observability for secure scaling.
Agentic patterns reduce engineering for massive business impact.
GoML is all in on AWS AI
These sessions confirm AWS is going to be the big fish in the agentic era. Cost effective models. Massive scale. AI-native data. Simplified storage. Autonomous architectures.
From GoML’s perspective as a leading AWS Generative AI development company, the announcements from Garman including Nova 2, the expanded Bedrock model lineup, and Lambda Durable Functions significantly strengthen what’s possible with agentic AI. These capabilities help drive measurable outcomes such as reducing clinician burnout through unified patient insights, accelerating pharmaceutical compliance processes to days instead of weeks, generating millions in additional asset revenue, and enabling near-instant fraud detection.
Together, these advancements allow GoML to deliver meaningful ROI more quickly for customers. Organizations looking to build their competitive edge with AWS-powered AI can learn more about our AWS AI services.
If you missed it, our recap has you covered for AWS re:Invent Day 1.
