Amazon rethinks its AI strategy and winds down many in-house models

Amazon Adjusts AI Development Priorities, Deprecates Several Nova Models

Amazon is reorganizing elements of its artificial intelligence efforts, including the deprecation of multiple in-house Nova models, adjustments to team structures, and a concentration of resources on a narrower set of frontier model projects, according to people familiar with the matter.

The moves follow staff reductions in the company’s Artificial General Intelligence organization last week and the closure of AGI Lab, a research unit established in 2024 after Amazon hired most of the team from AI startup Adept. The changes indicate a shift in resource allocation rather than a complete withdrawal from prior model lines.

Amazon has begun phasing out several Nova models, including the Premier and Omni variants, the Reel video-generation model, and the Canvas image-generation model. Some employees have referred to these models as operating in “KTLO,” an internal term indicating continued support for existing customers without active major development investment. The remaining Nova offerings include the Nova 2 Sonic and Nova 2 Lite foundation models, the Nova Forge customization service, and Nova Act agent technology.

Resources have been redirected toward a new initiative known internally as Frontier Model Research, or FMR, led by researcher Pieter Abbeel, who joined Amazon through the acquisition of robotics startup Covariant. The group is developing a new foundation model expected to be introduced at the company’s re:Invent conference later this year. An Amazon spokesperson stated that the company continues to support production models for customers and remains committed to frontier model work, noting that model portfolios are regularly updated based on customer requirements with migration paths provided.

The organizational adjustments do not eliminate the Nova brand entirely. The new model under development could be released under that name. The broader changes reflect a series of shifts that began after the departure of Rohit Prasad, who led the AGI organization until December 2025. Amazon subsequently placed the AGI group under senior vice president Peter DeSantis, combining it with silicon development and quantum computing responsibilities.

DeSantis has directed engineering and computing resources toward fewer frontier-model projects compared with the previous approach, which encompassed multiple text, image, and video model families. Employees have reported limited information on the long-term status of remaining Nova models. The recent staff reductions affected researchers who had been viewed as high-priority technical talent, and the AGI organization had maintained separate leveling and compensation structures to attract AI specialists.

Earlier public statements from the company had highlighted different priorities. At the 2025 re:Invent conference, AWS introduced Nova Omni 2 as a flagship multimodal model. In 2023, Chief Executive Andy Jassy had emphasized the AGI organization’s role in building ambitious foundation models, leading to the creation of six research groups. The current adjustments represent a narrower focus within less than three years.

An Amazon spokesperson said the company has historically maintained AI models in production for extended periods due to customer reliance and continues to evolve its lineup accordingly. The spokesperson added that clear guidance and migration options are provided as models advance.

The Frontier Model Research effort operates within the restructured AGI organization and has become its primary area of activity following the leadership transition. The new model under development is intended to address frontier capabilities, though specific performance targets or release details beyond the planned re:Invent timing have not been disclosed publicly.

Amazon’s approach to AI model development has involved both internal creation and acquisitions. The Adept-related AGI Lab focused on longer-term research before its closure. The Covariant acquisition brought Abbeel’s team into the Frontier Model Research group. These moves occurred alongside the company’s broader cloud and infrastructure investments through AWS.

Employees have described uncertainty regarding future project directions after the recent changes. The company has stated that such portfolio adjustments are standard practice in AI development and do not alter its overall commitment to the technology area.

The reported reorganization aligns with patterns observed at other large technology firms managing competing demands for specialized hardware and engineering talent in AI. Amazon has not released official statements detailing the exact number of positions affected or the full scope of model deprecations beyond the descriptions provided by people familiar with internal plans.