Amazon Web Services Inc. Chief Executive Matt Garman’s keynote at AWS re:Invent was filled with product updates with vision sprinkled in to help customers understand why the innovation matters.

To no surprise, this year’s keynote had a strong focus on the explosion of artificial intelligence and agents. The presentation outlined AWS’ strategy for empowering customers and developers in this new era by focusing on foundational infrastructure, diverse model choice, deep data integration, agentic platforms and developer tooling.

Here are five central themes from Garman’s keynote:

The inflection point: From AI assistants to billions of agents

Garman kicked off the keynote talking about how we are entering the era of billions of agents all working together to change the way we work and live. He declared that the advent of AI agents represents a major inflection point in AI’s trajectory, moving it from a “technical wonder” to a source of material business value that will be as impactful as the internet or the cloud. As he worked through his keynote, Garman discussed:

• Agents automating tasks for business value: Agents can perform tasks and automate actions on behalf of users, scaling people’s productivity up by 10x in some cases, leading to significant business returns across industries such as healthcare, customer service and payroll.
• Bedrock AgentCore scales agentic agents: The launch of Amazon Bedrock AgentCore is critical to empowering customers to deploy and operate highly capable agents securely at enterprise scale. AgentCore was built to open and modular, supporting frameworks like LangChain and models from various providers, including OpenAI and Gemini.
• Policy and Evaluations for trust: To address the need for predictability and control in autonomous agents (“trust but verify”), AWS introduced two key capabilities for AgentCore:
  • Policy in AgentCore: Provides real-time, deterministic controls over the specific actions agents can take with tools and data, using the Cedar open-source language to enforce boundaries.
  • AgentCore Evaluations: A new service to continuously inspect the quality of agent behavior (for example, correctness, helpfulness, harmfulness) based on real-world actions, automating a previously complex, data-scientist-heavy task.

AI infrastructure at planetary scale

A foundational theme of the keynote was the absolute necessity of highly scalable, secure and performant infrastructure to power the next generation of AI and agents. Garman emphasized that delivering the best AI performance and cost efficiency requires end-to-end optimization across hardware and software, a feat that AWS is uniquely positioned to achieve.

• Industry leadership in scale: Garman highlighted how AWS has the largest and most broadly deployed cloud infrastructure globally, with 38 regions and 20 Availability Zones. The sheer scale is underscored by the addition of 3.8 gigawatts of data center capacity in the last year, more than anyone else, and a private network that has grown 50%, to more than 9 million kilometers of cable.
• Purpose-built AI silicon: AWS continues to push the boundaries of price-performance with its custom-designed AI processors:
  • AWS Trainium: Although Tranium was built for training, Garman noted that Trainium 2 is excellent for inferencing and powers the majority of inferencing in Amazon Bedrock.
  • Trainium 3 Launch: The announcement of Trainium 3 UltraServers, featuring the very first three-nanometer AI chip in the AWS cloud, delivering 4.4 times more compute and five times more AI tokens per megawatt of power compared with its predecessor.
  • Trainium 4 Sneak Peek: A look ahead at the next iteration of the silicon, promising six times the FP4 compute performance.
• Best-in-class GPU experience: Garman stressed that AWS is the best place to run Nvidia graphics processing units, highlighting the operational stability and reliability achieved through 15-plus years of collaboration and sweating the small details (such as debugging BIOS) to avoid node failures. The launch of the new P6e GB300 instances, powered by Nvidia’s latest GB300 NVL72 systems, further supports this commitment.
• AWS AI Factories: AI factories have been all the rage, but they have typically been deployed on-premises with a hefty price tag. AWS’ versions bring more efficient pricing with stringent compliance and sovereignty requirements, allowing customers to deploy dedicated AI infrastructure that operates like a private AWS region within their own data centers.

Empowering choice and innovation with Amazon Nova and Bedrock

The belief that there will never be one model to rule them all is a core philosophy driving AWS’ model strategy, which is executed through Amazon Bedrock, the platform for generative AI applications.

• Model diversity: Bedrock continues to rapidly expand its selection, nearly doubling the number of models offered, including open-weights models like Google’s Gemma, Mistral Large, and Mistral 3, alongside proprietary models.
• Introducing Nova 2 family: Garman announced the new generation of Amazon’s own foundation models, Nova 2, designed to deliver cost-optimized, low-latency models with frontier-level intelligence. Nova 2 Light is a fast and cost-effective reasoning model, excelling at instruction following, tool calling and code generation. Nova 2 Pro is a more intelligent reasoning model for highly complex workloads, shining in areas critical for agents. Nova 2 Sonic is a speech-to-speech model for real-time conversational AI.
• Nova 2 Omni: A reasoning model that supports text, image, video and audio input and supports text and image generation output, addressing the need to understand multiple modalities simultaneously for real-world complexity.

Open training models with Nova Forge

Garman stressed that for AI to deliver value, it must be able to deeply understand a company’s unique data and intellectual property. To help customers with this he introduced open training models with Amazon Nova Forge.

• The data differentiator: The ability of models to understand a company’s data is what differentiates a business. Traditional methods like RAG and fine-tuning models on new domain data often hit limits, as models can “forget” core reasoning when customized post-training.
• Amazon Nova Forge: This new service gives customers access to a variety of Nova training checkpoints. Customers can blend their own proprietary data with an Amazon-curated data set at every stage of the model training.
• Creating Novellas: The result of this process is a “Novella” — a proprietary model that deeply understands the customer’s domain information without losing the foundational capabilities (such as reasoning) from the original training, enabling highly specific and intelligent guidance.

Reinventing how builders work

Finally, the keynote framed AI as a force multiplier for developers and enterprise teams, not just for enduser experiences. Innovations include:

• Amazon Q: This is a consumer AI experience for the enterprise that securely brings together all structured and unstructured company data (business intelligence data, databases, apps such as Microsoft 365 and more) to empower employees with research capabilities, BI insights and “quick flows” to automate repetitive tasks.
• Amazon Connect: AWS continues to lead in the contact center space with Amazon Connect, now a billion-dollar annualized business, pioneering AI-powered self-service and AI-driven recommendations for human agents. Amazon Connect was once a dark horse in an industry filled with legacy vendors, but that’s not the case anymore.
• AWS Transform: To free up developers to innovate, AWS Transform is an agentic tool focused on modernization. The new AWS Transform Custom allows customers to create custom code transformation agents to modernize any code, API, framework or language, even proprietary ones (for example, converting VBA to Python, Angular to React).
• Kiro: the agentic development environment: The primary developer agent is Kiro, an agentic development environment for structured AI coding. It has already been overwhelmingly adopted and standardized across Amazon for internal use.

Final thoughts

If one considers AWS a bellwether for AI, then Garman’s keynote can be considered a declaration that the cloud is now an AI-native platform, built from the ground up to empower a new era of autonomous invention. The core strategy is based on vertical integration, ensuring that foundational infrastructure can reliably scale the coming wave of AI agents.

The central theme for AWS in 2026 will be the mass enterprise adoption of autonomous AI agents, driven by the new capabilities of Bedrock AgentCore and the transformative efficiency gains promised by Kiro and Nova Forge’s custom model creation.

Zeus Kerravala is a principal analyst at ZK Research, a division of Kerravala Consulting. He wrote this article for SiliconANGLE.

Photo: AWS livestream