亚马逊加速推出 3nm 制程的 Trainium3，加速器已在少量数据中心安装并自周二起向客户开放，距上一代发布仅约一年，符合 Nvidia 每年一款新芯片的节奏。AWS 试图通过更优“价格—性能比”挑战 Nvidia 与 Google，强调 Trainium 在大规模模型训练中的成本效率。公司股价在消息公布后上涨 1.6%，而竞争对手 AMD 盘中下跌。当前劣势在于亚马逊芯片缺乏 Nvidia 的成熟软件生态，使部分用户（如 Bedrock Robotics）在关键模型训练上仍选择 Nvidia。
Trainium 的最大用户是 Anthropic，目前在印第安纳、密西西比与宾夕法尼亚等地的数据中心部署；AWS 已为其连接超过 50 万颗 Trainium，并计划年底前为 Anthropic 提供 100 万颗。尽管如此，亚马逊公开披露的其他大型客户有限，导致外界难以评估 Trainium 实际竞争力。Anthropic 同时也使用 Google TPU，并与 Google 达成获得“数百亿美元级算力”的协议，使其呈现多供应商策略。AWS CEO Matt Garman 称与 Anthropic 的关系“极其稳固”，并强调后者需求巨大。
亚马逊在 re:Invent 同步发布 Nova 2 模型系列，包括可处理文本、图像、语音与视频输入的 Omni，以及允许客户在训练未完成前接管与定制模型的 Nova Forge。此前 Nova 在行业基准中并非领先者，亚马逊强调“真实场景”表现优先。Nova Forge 已被 Reddit 用于定制违规内容识别模型，体现客户倾向选择专用模型而非最高端通用模型的趋势。
Amazon accelerated the rollout of its 3nm Trainium3 accelerator, now installed in a handful of data centers and available to customers beginning Tuesday, marking roughly one year since the prior Trainium release and matching Nvidia’s annual chip cadence. AWS aims to compete on price-performance, positioning Trainium as a cheaper and more efficient option for large-scale AI training. Amazon shares rose 1.6% after the announcement, while AMD fell intraday. A key drawback remains Trainium’s weaker software ecosystem relative to Nvidia’s, leading users such as Bedrock Robotics to rely on Nvidia hardware for model development.
Trainium’s largest user is Anthropic, with deployments across data centers in Indiana, Mississippi, and Pennsylvania. AWS has connected more than 500,000 Trainium chips for the startup and plans to dedicate 1 million by year-end. Few other major customers have been identified, leaving analysts uncertain about Trainium’s broader competitiveness. Anthropic also uses Google TPUs and secured access to tens of billions of dollars in Google compute, reinforcing a multi-provider strategy. AWS CEO Matt Garman described the Anthropic partnership as “incredibly strong,” citing the firm’s vast compute demand.
At re:Invent, Amazon also released its Nova 2 model family, including Omni, capable of accepting text, image, speech, or video inputs and producing text or image outputs. The company introduced Nova Forge, allowing customers to customize models before training concludes. Previous Nova models have not led industry benchmarks, with Amazon emphasizing real-world performance instead. Reddit is using Nova Forge to build a model that evaluates policy-violating posts, reflecting customer preference for domain-specific rather than maximal-scale general models.