亚马逊发布 Nova 模型家族

2024 年 12 月，亚马逊在 re:Invent 大会上发布了 Nova 系列基础模型。对关注云计算的人而言，这是一块等了很久的拼图——当微软有 OpenAI、谷歌有 Gemini、甚至阿里有通义千问时，亚马逊作为全球最大的云厂商，却长期依赖 Anthropic 等合作伙伴提供模型能力。Nova 的发布，标志着亚马逊终于补上了自研大模型这一课。

Nova 系列的定位很亚马逊：务实、性价比导向。它覆盖文本理解、创意生成、多模态与视频生成等能力，但宣传重点不是「最强」，而是「高效」——在保证质量的同时把推理成本压到合理区间。模型通过 Bedrock 统一对外，企业用户可以在自研 Nova 与 Anthropic Claude 等外部模型之间按需选择。

这种「自研 + 合作」的双轨策略，是亚马逊大模型战略的核心。一方面，自研 Nova 让亚马逊不必在模型定价与能力上完全受制于人；另一方面，继续与 Anthropic 深度合作（包括巨额投资与算力合作），让 AWS 生态始终拥有前沿模型供给。两条腿走路，既稳住了客户，也留住了选择权。

Nova 的意义超出了亚马逊一家公司。它再次证明，大模型已经从一个「可选的 AI 能力」变成了云厂商的「基础设施标配」——就像存储和计算一样，没有自己的模型，云平台就少了一块地基。AWS 的跟进，让云厂商自研模型的竞赛进入了最后补齐阶段。

回看 Nova 的发布，它最值得玩味的是时机与姿态。不抢首发、不喊最强，但在企业最需要性价比模型的时候稳稳落地。亚马逊用 Nova 完成了一次低调而重要的补齐——从此，AWS 生态里的每一家企业客户，都能在「自研」与「合作」之间，找到最适合自己的那款模型。

In December 2024 Amazon unveiled the Nova foundation models at re:Invent. For anyone following cloud computing, this was a long-awaited piece of the puzzle—while Microsoft had OpenAI, Google had Gemini, and even Alibaba had Qianwen, Amazon as the world's largest cloud vendor had long relied on partners like Anthropic for model capability. Nova marked Amazon finally filling the in-house model gap.

Nova's positioning was very Amazon: pragmatic, price-performance focused. It covers text understanding, creative generation, multimodal, and video generation, but the message was not "strongest" but "efficient"—keeping quality while pushing inference costs into a reasonable range. The models are offered through Bedrock, letting enterprise users choose between in-house Nova and external models like Anthropic's Claude as needed.

This "self-built plus partner" dual track is the core of Amazon's model strategy. On one hand, in-house Nova means Amazon is not fully at others' mercy on model pricing and capability; on the other, deep cooperation with Anthropic (including major investment and compute partnership) keeps frontier models in the AWS ecosystem. Walking on two legs holds customers steady and preserves choice.

Nova's significance reaches beyond one company. It again proved that large models have moved from an "optional AI capability" to "standard cloud infrastructure"—like storage and compute, without your own model, a cloud platform lacks part of its foundation. AWS's entry pushed the cloud-vendor in-house race into its final catch-up phase.

Looking back at Nova's release, the most interesting parts are timing and posture. No rush to be first, no claim of being strongest, but a steady landing exactly when enterprises needed price-performance models. Amazon used Nova to complete a low-key yet important gap-fill—from now on, every enterprise customer in the AWS ecosystem can find the most suitable model between "self-built" and "partner."