Las Vegas, November 30 to December 4, 2026. Caesars Forum, MGM Grand, The Venetian, Wynn, Encore, and Caesars Palace.
There is a version of the AWS re:Invent story that is about scale. The six-venue footprint spanning Las Vegas hotel corridors. The 2,000-plus sessions and workshops. The private fiber network running more than 9 million kilometers that AWS cites as the backbone of its global infrastructure. The 3.8 gigawatts of data center capacity added in a single year. The 38 regions and 120 availability zones that make AWS the most geographically distributed commercial computing platform ever built.
That version of the story is true and not particularly interesting.
The more useful version of the story in 2026 is about what AWS is actually doing with all of that infrastructure, and the answer is increasingly specific: building the production environment for AI agents at scale. Not AI as a feature, not AI as a capability layer, but AI agents as the new unit of computation that runs on top of the cloud the way applications once ran on top of operating systems.
This is the frame that Matt Garman, AWS CEO, has been developing through successive re:Invents, and it shapes how the 2026 program is organized, what the announcements emphasize, and what the hands-on labs are designed to teach. For the tens of thousands of developers, architects, and technology executives who descend on Las Vegas every December, understanding this frame is the prerequisite for understanding what AWS is building and why the decisions made at re:Invent 2026 will shape enterprise computing for the next several years.
The Agent Bet and What It Actually Means
Garman’s argument, developed through the 2025 re:Invent keynote and extended at 2026, is that AI agents represent the same kind of inflection point that the internet and the cloud each represented in their respective eras. Not incremental improvement, but a change in the fundamental unit of productive digital work.
The internet changed what information was accessible. The cloud changed what computing resources were available. Agents change what work can be automated without requiring humans to specify every step.
That third change is the most consequential because it removes the bottleneck that has limited software automation since software existed: the requirement that a human programmer specify the logic for every case the software needs to handle. An agent that can reason about a situation and determine appropriate action, rather than executing a predetermined workflow, can address the enormous class of business processes that are too variable, too context-dependent, and too judgment-intensive for conventional automation.
The practical implications are visible in the AWS product announcements that have accumulated over the past two years. Amazon Bedrock is the managed service through which customers access foundation models, but the significant development is not the models themselves. It is AgentCore, the infrastructure that makes it possible to run agents reliably in production: with deterministic policy controls over what tools and data agents can access, with continuous evaluation of agent behavior for accuracy and safety, and with the ability to connect agents to multiple model providers rather than locking into a single foundation model.
The Cedar language, which underlies AgentCore’s Policy feature, is the technical foundation for the trust problem that makes agent deployment difficult in practice. When a human employee makes a decision that turns out to be wrong, there is a clear accountability chain and a correction mechanism. When an AI agent makes a decision that turns out to be wrong, the question of what went wrong and why is harder to answer, and the question of what guardrails should have been in place is harder to specify after the fact. Cedar provides a declarative way to specify those guardrails before agents go into production, which is the prerequisite for organizations that need to manage legal and operational risk in agent-assisted workflows.
The Nova 2 Family and What It Tells You About AWS’s Model Strategy
The Amazon Nova 2 model family, which anchors AWS’s foundation model offerings at re:Invent 2026, reflects a specific strategic bet about how the foundation model market will develop.
AWS is not betting that one model will dominate. Garman has been explicit about this, and it is one of the clearer instances of a technology executive making a prediction that is actually falsifiable rather than safely vague. The bet is that different workloads will require different models, that open-weight models will capture significant market share alongside proprietary ones, and that the platform that makes it easiest to use any model for any workload will be more durable than any specific model investment.
Nova 2 Light is optimized for cost-effective reasoning: the workloads where inference cost matters more than maximum capability, which is a large fraction of actual enterprise AI deployments. Nova 2 Pro handles complex workloads where reasoning quality is the binding constraint. Nova Sonic is the speech-to-speech offering that supports the voice interface layer increasingly required for customer-facing applications. Nova Omni is the multimodal variant that handles combined text, image, and audio input.
Nova Forge, which allows customers to blend their own data with Amazon-curated data during training to produce custom models, is the product that addresses the gap between off-the-shelf foundation models and fully custom models. Most enterprises do not have the data volume or training infrastructure to train a competitive foundation model from scratch. But many have enough proprietary data to meaningfully differentiate a custom model from the base model for their specific domain. Nova Forge is the commercial path for that differentiation without requiring customers to build the training infrastructure themselves.
The near-doubling of models available in Bedrock, with the addition of open-weight models including Gemma and Mistral alongside the proprietary offerings, reflects the multi-model strategy in practice. Bedrock’s value proposition to enterprise customers is not superior models. It is a governed environment where any model can be deployed with consistent security, compliance, and cost management, without requiring a separate operational stack for each model provider.
The Infrastructure Investment That Makes the Agent Bet Possible
The infrastructure announcements at re:Invent 2026 are where the agent strategy meets the physical reality of what running AI agents at enterprise scale actually requires.
Trainium 2 currently handles most Bedrock inference. Trainium 3, built on 3nm silicon, delivers the next generation of inference performance. Trainium 4 is in preview, extending the roadmap visibility that infrastructure-dependent enterprises need to plan multi-year workloads. The NVIDIA P6e GB300 instances reflect AWS’s commitment to offering customers the choice of AWS custom silicon and NVIDIA GPU infrastructure rather than forcing a choice between them.
AWS AI Factories, which deploy dedicated AI infrastructure inside customer data centers in a configuration that functions like a private AWS region, address the specific requirements of customers who cannot send sensitive data to public cloud infrastructure for regulatory, sovereignty, or competitive reasons. The AI Factory model preserves the operational model that makes AWS valuable, managed infrastructure, consistent tooling, centralized visibility, while satisfying the data residency and control requirements that prevent some workloads from moving to the public cloud at all.
The 38-region, 120-availability-zone footprint is not a legacy of the pre-AI era. It is the geographic distribution that makes agent deployment at enterprise scale possible for global organizations whose customers, data, and regulatory requirements are distributed across jurisdictions with different rules about data processing and AI deployment. An agent serving a financial institution’s European customers needs to process data in Europe. An agent supporting a healthcare system’s patient data needs to stay within healthcare data governance boundaries. The geographic infrastructure is the enabling layer for the policy controls that make compliant agent deployment possible.
Amazon Q and the Enterprise Knowledge Layer
Amazon Q, AWS’s enterprise AI assistant, represents a different approach to the enterprise AI opportunity than Bedrock’s infrastructure-first model.
Where Bedrock provides the building blocks that developers use to build AI applications, Amazon Q is a completed application that enterprise knowledge workers use directly, without requiring development. It connects to enterprise data sources through standard connectors, builds a knowledge graph of the organization’s information environment, and responds to questions and tasks in a way that is grounded in that enterprise context rather than in the general training of the foundation model alone.
The reported billion-dollar annualized business that Amazon Connect has become reflects a parallel dynamic at the customer service infrastructure layer: AI-augmented contact center operations at enterprise scale, where the value comes from the combination of foundation model reasoning and enterprise-specific context rather than from either alone.
AWS Transform, which offers custom modernization agents that help enterprises migrate legacy applications and infrastructure to modern AWS equivalents, is the product that makes the agent strategy concrete for the large enterprises that have the most to gain from AI-assisted modernization and the most resistance to the disruption it requires. The legacy application modernization problem has been one of the largest sources of technical debt in enterprise computing for decades. Agents that can analyze, document, and migrate legacy codebases without requiring the traditional consulting-intensive approach represent a genuine change in the economics of that problem.
Kiro, the agentic development environment that AWS announced at re:Invent, is the product aimed at developers rather than enterprise buyers. It is an IDE-level agent that assists with code generation, review, debugging, and documentation in the context of the developer’s actual codebase rather than in the abstract. The development environment has been the space where most AI coding tools have competed, and Kiro’s integration with the broader AWS toolchain gives it a natural advantage in environments where developers are already building on AWS infrastructure.
What 2,000 Sessions Actually Contain
The re:Invent session structure, at 2,000-plus sessions and workshops across five days, requires more deliberate planning than most attendees arrive with. The categories are worth understanding before attempting to navigate the schedule.
Keynotes are the highest-profile events and the ones most likely to carry the major announcements. Matt Garman’s CEO keynote typically anchors the Tuesday of the week. The technical keynote, usually focused on the services and infrastructure announcements that the developer community cares most about, often runs separately and rewards closer attention from technical practitioners than the main keynote.
Breakout sessions are the 60-minute presentations that cover specific services, architectures, or customer use cases. The best ones are practitioner-led, presenting real implementations with real data on what worked and what did not. The worst are thinly disguised product demos with case study framing.
Workshops are where the conference’s hands-on value concentrates. A three-hour workshop on building a production-ready agent with AgentCore, or on configuring Bedrock’s multi-model governance, produces a different kind of learning than any passive session format: the kind that transfers to actual work rather than informing awareness.
Chalk Talks are the smallest-format sessions, capped at 50 attendees, where practitioners can engage directly with AWS subject matter experts on specific technical questions. For attendees who arrive with specific architectural challenges rather than general orientation needs, Chalk Talks are the most valuable content at the conference.
Builder Sessions are the hands-on equivalent of Chalk Talks: small groups, hands-on with actual services, with AWS engineers available for real-time guidance. Registration fills months before the conference, and showing up without a reservation to a Builder Session is the most reliable way to be turned away from the most valuable content on the floor.
The Competitive Context That Makes 2026 Significant
AWS re:Invent 2026 is happening at a moment when the competitive dynamics of the enterprise cloud market are more interesting than they have been since the hyperscaler era began.
Microsoft’s Azure has built a significant position on the strength of its OpenAI integration and the deep enterprise relationships that come from its Office and Teams presence. Google Cloud has made serious progress in AI infrastructure and in specific enterprise verticals where its AI research capability produces differentiated products. Oracle’s investment in AI infrastructure, including the Zettascale10 supercomputer and OCI Multicloud Universal Credits, represents a genuinely competitive alternative for specific workloads and customer profiles.
AWS’s response to this competitive environment at re:Invent 2026 is not to claim superiority across every dimension. It is to make the case for the breadth and maturity of the AWS ecosystem, the managed infrastructure model that reduces operational burden on enterprise IT teams, and the agent platform that is designed to be the production environment for the AI work that enterprises are beginning to move from pilot to deployment.
The OCI Multicloud Universal Credits story, which runs in parallel with the Oracle AI World announcements, is worth tracking from the AWS perspective. Cross-cloud workloads that span AWS and OCI create architectural questions that neither company’s conference addresses fully, and the practitioners at re:Invent who are managing multi-cloud environments will find that the AWS sessions do not always provide the context required for optimal cross-cloud design decisions.
Key Takeaways
AWS re:Invent 2026’s organizing frame is agents: AI agents as the new unit of production computation, AgentCore as the governance and operational infrastructure for deploying them reliably, and the full AWS ecosystem as the platform on which they run. Understanding this frame is the prerequisite for understanding why the specific announcements, Nova 2, Nova Forge, AgentCore, Kiro, AI Factories, are structured the way they are.
The multi-model strategy embedded in Bedrock, including near-doubled model availability with both open-weight and proprietary options, reflects a genuine architectural choice about how to position a cloud platform in a market where no single foundation model is likely to dominate all use cases. Customers who agree with this assessment should evaluate Bedrock’s model governance and operational consistency rather than model capability alone.
AgentCore’s Policy feature and the Cedar language are the trust infrastructure that makes enterprise agent deployment possible at scale. Organizations that are serious about moving agent pilots to production should evaluate AgentCore’s governance model before building custom agent orchestration, because the policy specification problem is harder than it looks and Cedar provides a more tractable solution than hand-coded guardrails.
The AI Factory deployment model addresses a real gap in the enterprise AI adoption landscape: the organizations that need AI at scale but cannot send their most sensitive workloads to the public cloud. For regulated industries, the AI Factory model may be the path to agent deployment that was previously unavailable.
Builder Sessions and Chalk Talks are the highest-value content at re:Invent for practitioners with specific implementation challenges. Registration priority should go to these formats before breakout sessions and before general access to the expo floor.
Notes for Technology Leaders and Practitioners
Conference pass registration for re:Invent opens months before the November 30 start date. The $2,499 full conference pass is the appropriate tier for practitioners who will attend workshops and need access to the full session library. The early bird pricing at $1,299, which closed August 25, is the signal that advance planning pays off significantly in cost as well as in session availability.
Builder Session and Chalk Talk registration is separate from the conference pass and fills faster. The most sought-after technical sessions on agent architecture, Bedrock configuration, and Trainium optimization require reservation plans that should be developed as soon as the session catalog publishes.
The Venetian and Caesars Forum are typically the densest venues for AI and infrastructure content. MGM Grand and Wynn tend to concentrate enterprise and business-track sessions. Understanding the venue geography reduces the time spent navigating between sessions in a conference sprawl that spans multiple casino properties connected by shuttle.
The expo floor, open across the conference days, carries both AWS service demonstrations and the third-party partner ecosystem. Partner announcements at re:Invent often receive less visibility than the AWS service announcements but can represent significant capability additions, particularly for practitioners working in specific vertical markets where ISV tools add the domain specificity that the generic AWS services do not provide.
For practitioners evaluating the multi-cloud implications of their AI strategy: the SuiteWorld 2026 announcements running concurrently, including NetSuite Next’s AI Connector Service on MCP and Oracle AI World’s Zettascale10 and Multicloud Universal Credits, represent the competitive context that shapes AWS’s 2026 positioning. Understanding what the competition is doing is the prerequisite for evaluating what AWS is offering.
AWS re:Invent 2026 runs November 30 through December 4 across six Las Vegas venues: Caesars Forum, Caesars Palace, Encore, MGM Grand, The Venetian, and Wynn. For registration, session catalog, and partner information, visit aws.amazon.com/reinvent.
