Onome Amuge
Amazon Web Services (AWS) has unveiled a series of next-generation computing and artificial intelligence offerings designed to accelerate enterprise transformation and broaden access to AI capabilities, signaling a shift in the competitive dynamics of cloud infrastructure. At its annual re:Invent 2025 conference, the company highlighted innovations that combine performance, efficiency, and scalability, reflecting AWS’s growing emphasis on lowering the barriers to AI adoption for organizations of all sizes.
Central to the announcements was the introduction of Graviton5, AWS’s most powerful and efficient CPU to date. Built with 192 cores and a fivefold increase in cache over its predecessor, the Graviton5-based Amazon EC2 M9g instances promise up to 25 per cent higher performance than the previous generation. AWS reports that for the third consecutive year, more than half of new CPU capacity deployed across its platform is powered by Graviton, with 98 per cent of its top 1,000 EC2 customers, including Adobe, Airbnb, Epic Games, Formula 1, Pinterest, SAP, and Siemens, already benefiting from the chip’s combination of price and performance.
“The complexity of modern workloads requires not just raw power, but also energy efficiency and predictability at scale,” said an AWS spokesperson. Graviton5 exemplifies this approach, allowing organizations to meet performance and sustainability goals without trading off speed or cost-effectiveness.
AWS also expanded its Nova family of AI models, introducing four new offerings optimized for reasoning, multimodal processing, conversational AI, code generation, and agentic tasks. Notably, Nova Forge pioneers what AWS calls “open training,” giving enterprises access to pre-trained model checkpoints and enabling them to blend proprietary datasets with curated Amazon Nova data. Early adopters, including Reddit and Hertz, report accelerated development cycles, with Nova Act delivering 90 per cent reliability for browser-based UI automation and reducing iteration times fivefold.
A further innovation came in the form of three “frontier agents”,; autonomous AI agents that operate as extensions of enterprise teams. Kiro, the autonomous agent, functions as a virtual developer; the AWS Security Agent acts as a continuous security consultant; and the AWS DevOps Agent provides on-demand operational support. Commonwealth Bank of Australia, SmugMug, and Western Governors University are among early users integrating frontier agents into software development lifecycles, illustrating a potential paradigm shift in enterprise operations where AI assumes persistent, autonomous roles previously held by human teams.
On the hardware front, AWS introduced Trainium3 UltraServers, powered by the company’s first 3nm AI chip. Each server integrates up to 144 Trainium3 chips, delivering up to 4.4 times more compute performance and four times the energy efficiency of the prior generation. Early customers report halving training and inference costs while cutting model training timelines from months to weeks. Decart, for example, has achieved fourfold faster inference for real-time generative video at half the cost of traditional GPU setups. The company’s Bedrock platform is already leveraging Trainium3 for production workloads, highlighting the growing interplay between infrastructure and enterprise AI deployment.
For organizations constrained by regulatory or infrastructure considerations, AWS also unveiled AI Factories, a service that brings high-performance AI infrastructure directly into customer data centers. Combining NVIDIA GPUs, Trainium chips, and AWS networking with software services like Amazon Bedrock and SageMaker AI, AI Factories enables enterprises and governments to deploy AI without massive upfront capital outlays while maintaining data sovereignty and regulatory compliance.
Complementing the hardware and model expansions, AWS introduced enhancements to its AgentCore platform. New features include Policy, which allows teams to define natural-language boundaries for agent actions; Evaluations, providing 13 pre-built metrics for real-time performance monitoring; and episodic memory functionality, enabling agents to learn from prior interactions. Together, these developments underscore AWS’s ambition to make AI both operationally reliable and contextually intelligent.
