New chips, services for businesses and alliances with Ai bigwigs: all the news from Aws

'Everything at Amazon starts with the customer,' Garman relaunched to explain the strategic choice of accompanying companies, from large down to start-ups, in building agents, with interventions on the entire supply chain, from infrastructure to complete solutions, to enable the creation of AI agents on a large scale and in an affordable manner.

Thus, at the infrastructure level, the offering was enhanced by Ai Factories, to bring Aws services into customers' existing data centres, respecting sovereignty, policy and governance, for projects often in the public sector. At the same time, Garman announced the launch of the new Trainium 3 chip with twice the energy efficiency of version 3, targeted at large-scale training and global inference. And which is adopted by its partner of excellence Anthropic for the training of the new Claude generation. Although the long-standing collaboration with Nvidia continues, Aws has thus chosen to pursue the refinement of its own family of chips dedicated to training: the Trainium 3 marks a fivefold increase in output with the same power consumption, but work is already underway on version 4 with a performance target of six times more efficient.

Basically, at the inference level, Aws leverages Bedrock's global platform, which has doubled its base of generative models over the past year and now encompasses the entire global generative AI offering, with a focus on the latest version of the European champion Mistral. In this area, Aws has enriched Nova's family of foundational models: in addition to the fast reasoning of Lite, the advanced high-accuracy reasoning of Pro, and the real-time voice-to-voice connections of Sonic, it has now also added the multimodal model of Omni, which integrates text, images, video and voice, simplifying creative flows with one of the first tools for the natural orchestration of different models.

At the level of data management, the Nova Forge alternative is offered, an open training option that balances training and adaptation with the integration of proprietary data, aiming for the most customised result possible: 'Your data is unique: it's what differentiates you from your competitors,' Garman reiterated, emphasising the need to make the most of this valuable asset.

Finally, Aws aims to provide tools to concretely enable companies in the construction of their agents, with the ambition of accelerating time-to-value by providing, on the one hand, policies to define their boundaries of action without distorting the non-deterministic nature of their behaviour and, on the other hand, with a constant and automatic evaluation in terms of correctness, usefulness and security, with a continuous update that lightens the work of company teams.

The ultimate goal is therefore to put 'AI agents to work', simplifying the path for companies as much as possible. In this area, Aws continues to refine tools such as Quick Suite, a 'corporate companion' that allows everyone to move quickly from insight to action, or Amazon Connect, a workflow automation tool based on Ai-human interaction. But it also presented Transform, a tool capable of transforming any antiquated computer code or language, resulting in a reduction of technical debt from legacy systems. At the same time, the group introduced specific tools to support developers in programming and a class of autonomous agents to solve problems, again in the IT domain, in full autonomy.

Having lagged behind in the Large Language Models, Amazon Web Services has chosen to open its platform to the most advanced models in order to support customers in the simple adoption of customised and profiled solutions closer to their needs, feeding the value-added services chain closer to the enterprise, with autonomous, scalable and long-lasting solutions.