Amazon is consolidating its artificial intelligence research efforts, shuttering its AGI Lab and retiring most Nova models to concentrate resources on a single, high-capacity frontier system.
Amazon is fundamentally restructuring its artificial intelligence research operations, shifting from a diversified model development strategy to a singular focus on a high-capacity frontier system. This strategic pivot, reported by Business Insider, involves the retirement of the majority of the Nova model family and the closure of the AGI Lab as the organization reallocates its computational and human capital.
The transition marks a departure from the multi-model approach initiated in 2023 under the former AGI organization leadership. During that period, Amazon developed a broad suite of generative tools, including the Nova Omni 2 multimodal reasoning system, the Reel video generation tool, and the Canvas image synthesis framework.
Leadership changes within the division have catalyzed this consolidation. Following the departure of Rohit Prasad in December 2025, the AGI organization was integrated into the portfolio of Peter DeSantis, senior vice-president of silicon and quantum computing. This structural shift has resulted in a narrower strategic mandate focused on frontier-scale research rather than the maintenance of disparate model families.
Engineering teams have transitioned legacy Nova projects into a maintenance-only state, colloquially referred to as keep-the-lights-on. While systems like Nova 2 Sonic, Nova 2 Lite, and the Nova Forge platform remain operational for existing enterprise customers, active development cycles for the broader Nova portfolio have ceased.
The Frontier Model Research group receives the majority of the reallocated resources under the direction of researcher Pieter Abbeel. Abbeel, who joined Amazon following the acquisition of the robotics firm Covariant, is overseeing the development of a new flagship foundation model. This system is currently slated for a potential unveiling at the upcoming re:Invent conference, though its final branding remains subject to internal review.
The restructuring has impacted the organization’s specialized research infrastructure, including the shuttering of the AGI Lab. This facility, established in 2024 following the integration of technology and personnel from the startup Adept, previously functioned as a hub for long-term exploratory research. The dissolution of the lab and associated layoffs reflect a broader trend of prioritizing immediate, high-impact model training over decentralized experimental research.
The internal reallocation of compute resources is significant given the specialized nature of the AGI organization. Previously, this group operated with its own distinct levelling and compensation structures, specifically designed to attract top-tier AI engineering talent. By folding these operations into the broader silicon and quantum computing portfolio, Amazon is effectively subjecting its research division to the same rigorous resource-efficiency standards applied to its hardware and cloud infrastructure divisions.
Amazon abandoned the broad Nova ecosystem to prioritize the development of a competitive, state-of-the-art foundation model capable of rivaling existing frontier systems. This move forces a trade-off between product breadth and the depth of a single, highly optimized model architecture. The success of this pivot will likely be measured by the performance metrics of the upcoming foundation model against established benchmarks in reasoning and multimodal processing.
The consolidation of research efforts suggests a shift toward optimizing compute-to-parameter ratios for a single, massive-scale model. By concentrating its hardware resources, Amazon aims to overcome the latency and performance limitations inherent in managing multiple, smaller-scale model architectures. This strategy aligns with the industry-wide trend of scaling laws, where performance gains are increasingly tied to the sheer volume of compute allocated to a unified training run.
The decision to centralize research reflects the high cost of maintaining diverse model families in a competitive environment. Maintaining separate teams for text, image, and video generation requires redundant infrastructure and fragmented data pipelines. By focusing on a single frontier model, the company reduces overhead and concentrates its engineering talent on the most critical path for model performance.
Industry observers will monitor the upcoming re:Invent conference for technical disclosures regarding the architecture of the new frontier system. The transition period will test the company’s ability to maintain existing enterprise commitments while simultaneously executing a high-stakes migration of its core research talent toward a singular, unified training objective.
