AWS re:Invent 2025: Amazon Lays out Strategy for Agentic AI and Silicon Sovereignty

AWS re:Invent 2025 marks a decisive pivot from the generative chat interfaces that defined the last two years toward “Agentic AI” (autonomous software capable of executing complex workflows rather than merely generating text).

This strategic shift arrives as enterprises face mounting pressure to demonstrate returns on significant AI investments, moving beyond experimental pilots to production-grade systems.

Matt Garman, CEO of AWS, framed this evolution as a necessity for unlocking genuine business value. “AI assistants are starting to give way to AI agents that can perform tasks and automate on your behalf. This is where we’re starting to see material business returns from your AI investments.”

The Agentic Shift: From Chat to Action

Central to this strategy is the general availability of Amazon Nova Act, a specialized model designed specifically for browser automation.

While general-purpose large language models often struggle with interface navigation, Nova Act achieves a 90% success rate on tasks such as form filling, data extraction, and complex booking workflows, according to Amazon.

To achieve this reliability, the model training process moved beyond standard text corpora to simulate actual user workflows. Danilo Poccia, Chief Evangelist at AWS, explained the methodology in the official launch post:

“Nova Act addresses the challenge of building reliable browser automation at enterprise scale. Powered by a custom Amazon Nova 2 Lite model, Nova Act excels at driving browsers, support calling APIs, and escalating to humans when needed.”

“Nova Act approaches this differently by using reinforcement learning while the agents run inside custom synthetic environments (‘web gyms’) that simulate real-world UIs. This vertical integration across the model, orchestrator, tools, and SDK, all trained together, unlocks higher completion rates at scale.”

Deploying these capabilities happens through the AgentCore platform, which has been updated with enhanced policy controls. Organizations can now define strict boundaries for autonomous agents, preventing them from hallucinating actions or accessing restricted data systems.

Swami Sivasubramanian, AWS’s Vice President of Agentic AI, characterized the platform as a mechanism to remove friction from the development process. By offloading complex orchestration to autonomous agents, the company argues that developers can significantly reduce the time required to translate conceptual prototypes into functioning, high-impact production systems.

Silicon Sovereignty: Challenging Nvidia

Powering these autonomous agents requires significant computational resources, a reality AWS is addressing through aggressive vertical integration. Unveiled during the keynote, the Trainium3 (Trn3) UltraServers represent the first AWS AI chips built on a 3nm process node.

According to the technical specifications, the new Trn3 UltraServers maximize density by packing up to 144 chips into a single integrated system, a configuration that yields a 4.4x increase in compute performance over the previous Trainium2 generation.

Benchmarks running OpenAI’s open-weight GPT-OSS models suggest the architecture translates to three times higher throughput per chip and four times faster response times. Beyond raw speed, the architectural improvements also target operational costs, with the new chips delivering 40% better energy efficiency than their predecessors.

Directly challenging Nvidia’s market dominance, this hardware push offers a lower-cost alternative for heavy training workloads. Deployment is already underway at the large-scale Project Rainier data center in Indiana, which opened in October to support Anthropic’s model training.

On the CPU side, the new Graviton5 processor aims to maximize density for general-purpose cloud workloads, delivering a 25% performance improvement.

Architecturally, the Graviton5-based EC2 M9g instances focus on maximizing density and data throughput. By packing 192 cores into a single package, AWS claims the highest CPU core density currently available in its cloud fleet.

To prevent data bottlenecks in such a dense environment, the design includes a significantly expanded memory buffer, featuring a 5x larger L3 cache overall. On a per-core basis, this translates to 2.6 times more cache access compared to the Graviton4. The upgrade extends to I/O performance as well, with average increases of 15% for network bandwidth and 20% for Amazon Elastic Block Store (EBS) throughput.

The Developer Reality Check

While the infrastructure promises efficiency, the developer tools layer has sparked discussions this summer. Controversy sparked immediately regarding the general availability of the Kiro AI coding agent, which introduced a complex metering system distinguishing between “Vibe” (chat-based) and “Spec” (task-based) usage.

Despite the pricing friction, executives remain committed to the product’s long-term vision. Amazon CEO Andy Jassy stated back then, “Kiro has a chance to transform how developers build software.”

Early adopters, however, have criticized the pricing structure, noting that heavy usage could lead to monthly bills exceeding $1,900 per developer. Such friction highlights the broader industry challenge of balancing advanced AI capabilities with predictable operational costs.

Enterprise ROI and Market Context

Innovations land in a market becoming increasingly skeptical of AI hype. Analysts note that while infrastructure spending is booming, evidenced by the $38 billion cloud deal OpenAI signed with AWS in November, enterprise adoption is lagging behind the technology’s capabilities.

“AWS AI announcements show that AWS is thinking ahead and maybe far too ahead,” Chhabra wrote. “Most enterprises are still piloting AI projects and are rarely at the levels of maturity AWS expects them to be to take advantage of the offerings that come out of these announcements.”

Addressing cost concerns, AWS introduced Database Savings Plans, offering up to a 35% reduction in costs for committed usage across Aurora, RDS, and DynamoDB. Cloud economists, who have long argued for better cost controls, welcomed the move. Corey Quinn, Chief Cloud Economist at Duckbill, remarked, “Six years of complaining finally pays off.”

Infrastructure as a Service

Highly regulated industries unable to move fully to the public cloud can now utilize “AI Factories.” Governments and large enterprises can deploy fully managed AWS AI infrastructure, including Trainium chips and Bedrock services, within their own private data centers.

Market observers point to Amazon’s physical infrastructure as a key differentiator in this segment.

Ethan Feller, an Equity Strategist at Zacks Investment Research, views this strategy as a realignment with Amazon’s fundamental strengths. Rather than solely competing in the crowded model layer, he argues that AWS is capitalizing on its dominance in the underlying cloud infrastructure.

By focusing on “where the models are run,” Amazon is effectively positioning itself as the essential utility provider for the AI economy, leveraging its logistical expertise to capture value regardless of which specific AI model wins the market.

By controlling the entire stack, from the 3nm silicon to the agentic orchestration layer, AWS is betting it can force the market past the current “trough of disillusionment.”

Open questions remain regarding whether enterprises are ready to hand over control to autonomous agents. Werner Vogels, CTO of Amazon, addressed this anxiety directly: “Will AI take my job? Maybe. So maybe we should rephrase and reframe this question. Will AI make me obsolete? Absolutely not, if you evolve.”