Amazon's AI Strategy Shift: Abandoning 'Multi-Pronged' Approach, Focusing on New Model Development, and Strengthening 'Gateway Platform' Identity

[Ebrun Original] July 31 news - According to foreign media reports, Amazon is comprehensively restructuring its artificial intelligence strategy, gradually halting the advancement of multiple internally developed models, reorganizing relevant teams, and consolidating engineering resources toward new strategic directions to enhance the company's competitiveness in the frontier AI field. Instead of simultaneously investing in the development of multiple text, image, and video models as in the past, the company is now concentrating engineering talent and limited computing resources on higher-priority projects. Reports indicate that Amazon has begun deprioritizing several internal flagship Nova models, including the high-end models Premier and Omni, the video generation model Reel, and the image generation model Canvas. Some Amazon employees describe these models as being in a "maintenance mode": they still provide support for existing customers but are no longer a focus for R&D resource allocation. Meanwhile, Amazon's R&D efforts are shifting toward a new frontier model research initiative, internally dubbed "Frontier Model Research" (FMR), which has become one of the highest-priority projects in Amazon's AI strategy this year. Under this initiative, Amazon is developing a new flagship foundational model, expected to be unveiled at the annual developer conference re:Invent this fall. An Amazon spokesperson stated in an interview that AI models remain one of the company's most important ongoing efforts, and this has not changed; like any AI product portfolio, the company continuously adjusts its model lineup based on customer needs, always providing clear guidance and model upgrade migration paths for customers. However, this organizational change does not mean Amazon is completely abandoning the Nova brand. Currently, the Nova series still includes the Nova 2 Sonic and Nova 2 Lite foundational models, the Nova Forge service for building and customizing models, and Amazon's AI agent technology, Nova Act. In fact, the new model being developed under the FMR project may also continue to use the Nova brand in the future. This adjustment reflects Amazon's internal restructuring of its AI business over the past year. In 2023, Amazon established the AGI (Artificial General Intelligence) division, which includes multiple specialized teams, one of which is AGI Lab. Amazon founded this lab in 2024 to drive long-term AI research. In February of this year, the core figure of the team departed, and Amazon closed AGI Lab in the latest round of organizational adjustments last week. Replacing it and taking center stage is FMR, also a team under the AGI division, responsible for developing Amazon's next-generation frontier AI models. Concurrently, there have been personnel changes at the top leadership of the AGI division, with a new head driving a more focused AI strategy. Previously, Amazon simultaneously advanced multiple model series across text generation, image generation, and video generation, but the new strategy opts to concentrate engineering talent and computing resources on fewer, more cutting-edge model projects. This shift marks a significant change in Amazon's AI strategic direction. Just last year, AWS aggressively launched the Nova Omni 2 as a flagship multimodal reasoning model, with CEO Andy Jassy personally driving the construction of the AGI division, establishing six new research teams. Yet, only a year later, Amazon is abandoning the internal R&D approach of "attacking in all directions" with multimodal models. This decision is not without financial rationale. Amazon's latest earnings report highlights a fact: in the AI competition, a company does not necessarily need to possess the most powerful large model to succeed. In the second quarter, its cloud computing business AWS saw revenue grow 37% year-over-year, driven primarily by the cloud giant's ability to offer customers the capability to run multiple leading AI models, including those from Anthropic and OpenAI. Jassy explained that the "dominance window" for a single model to hold a technological lead is becoming shorter, with different models potentially surpassing each other at different stages, and each performing better on different tasks. Many customers want to use multiple leading models simultaneously rather than anchoring to a single model. This trend has fueled the rapid growth of Amazon's Bedrock platform, positioned as an "AI model supermarket aggregation gateway" infrastructure that enables businesses to access various foundational models. Jassy predicts that, within the next few years, the market will see at least "about six models with comparable performance." He noted that all these models will appear on the Bedrock platform, and one of them will be Amazon's own. In this regard, rather than solely betting on proprietary models to achieve industry leadership and win a performance competition, Amazon is now placing greater emphasis on making Bedrock the core of its AI strategy. The future role of self-developed models may be more as a bargaining chip to prioritize serving customer niche needs, reduce reliance on external suppliers, and optimize costs. Ebrun will continue to track and report on this intelligence. For more information related to this article, please scan the QR code to follow the author's WeChat.