Meta has entered into a large-scale strategic agreement with Amazon Web Services (AWS), under which millions of AWS Graviton processors are planned to be used to meet the social network's growing artificial intelligence needs. Representatives of Amazon announced this partnership on Friday, emphasizing a technological shift in approaches to building computational infrastructure. While graphics processing units (GPUs) remain an indispensable tool for the initial training of large language models, central processing units (CPUs) based on the ARM architecture, such as Graviton, are becoming critically important in subsequent stages.
After the resource-intensive model training stage is complete, developers' focus shifts to the operation of autonomous AI agents, which generate a fundamentally different type of computational load. Such systems require continuous power to perform everyday tasks in real time: generating program code, logical inference, complex searches, and coordinating actions when executing multi-stage requests. It is precisely for such use cases, known as inference, that the architecture of custom central processors is optimally suited.
The timing of this deal's public announcement highlights the high tension in the cloud services market. AWS management announced the signing of the contract immediately after the conclusion of the large-scale industry conference Google Cloud Next. This move appears particularly symbolic in the context of previous corporate decisions: last August, Meta signed a six-year contract with the Google Cloud platform for a total of ten billion dollars. Prior to this, the technology giant predominantly used server capacities from AWS and Microsoft Azure. The return of a significant portion of Meta's infrastructure budget back into the Amazon ecosystem represents an important strategic and reputational achievement for the company.
The successful application of its own microprocessors at the level of major corporate clients makes the cloud provider a direct competitor to traditional hardware manufacturers. Processors from the Graviton family serve as an alternative to solutions like NVIDIA's new Vera chip, which is also based on the ARM architecture and designed for agent workloads. However, the business models of the corporations differ fundamentally: while NVIDIA sells its physical chips and systems to enterprises and third-party cloud platforms, AWS provides access to its developments exclusively through its own services.
The conclusion of such large contracts places enormous responsibility on Amazon's internal chip development team, whose laboratories operate under maximum market pressure. The partnership with Meta organically complements the company's recent successes in promoting another proprietary development — Trainium AI accelerators, which are used for both training and inference. Earlier this month, Anthropic, the creator of the Claude model family, committed to spending one hundred billion dollars over ten years on using AWS infrastructure, with a particular emphasis on the Trainium architecture. In return, Amazon agreed to invest an additional five billion dollars in the startup, increasing its total capital investment in Anthropic to thirteen billion.
Sources
Replies (0)
No replies in this topic yet.
