AWS发射的金属芯片和超服务机,在扩大AI硬件竞争的同时,提高了性能、效率和降低成本。 AWS launched Trainium3 chips and UltraServers, boosting performance, efficiency, and lowering costs while expanding AI hardware competition.

亚马逊网络服务公司在re:Invent 2025展会上发布了其Trainium3 AI芯片和UltraServers,与之前的系统相比,它提供高达4.4倍的计算性能,4倍的能源效率,以及50%的成本降低. Amazon Web Services unveiled its Trainium3 AI chip and UltraServers at re:Invent 2025, offering up to 4.4x higher compute performance, 4x greater energy efficiency, and 50% lower costs compared to prior systems. 新芯片在每台144个芯片的服务器上使用,支持NVLink Fusion与Nvidia硬件无缝结合。 The new chips, used in servers with 144 chips each, support NVLink Fusion for seamless integration with Nvidia hardware. AWS还推出了AI工厂——受管制行业的房地基础设施——并更新了Nova AI模式,采用新福治等工具进行定制模式培训。 AWS also launched AI Factories—on-premises infrastructure for regulated industries—and updated its Nova AI models with tools like Nova Forge for custom model training. 该公司报告了像Anthropic这样的公司的强烈客户兴趣,并强调了内部效率的提高,同时扩大了其内部AI硬件战略,与Nvidia,Google和Meta竞争. The company reported strong client interest from firms like Anthropic and highlighted internal efficiency gains, while expanding its in-house AI hardware strategy to compete with Nvidia, Google, and Meta.