Amazon is streamlining its generative AI efforts by deprioritizing active development of several proprietary Nova models and concentrating engineering talent and computing resources on a narrower set of next-generation frontier model initiatives. This recalibration prioritizes high-impact research over maintaining a broad portfolio of specialized text, image, and video models. Existing models continue to receive support for customers already using them in production, with migration guidance provided as the lineup evolves. The shift aligns with broader industry patterns in which major technology firms narrow focus amid constrained specialized compute resources and intense competition at the frontier of model capabilities.
Background on the Nova Family
Amazon introduced the Nova family of foundation models at its AWS re:Invent conference in late 2024. The initial lineup included understanding models such as Nova Micro (text-only, optimized for low latency and cost), Nova Lite and Nova Pro (multimodal models handling text, images, and video), and the higher-capability Nova Premier for complex reasoning tasks. Complementary creative models included Nova Canvas for image generation and Nova Reel for video generation.
These models launched exclusively through Amazon Bedrock, AWS’s managed service for accessing foundation models. Amazon positioned Nova models around strong price-performance ratios, with advantages in speed and cost relative to comparable offerings in their intelligence classes, along with support for fine-tuning, distillation, and integration with enterprise data and systems. Subsequent expansions in 2025 brought Nova 2 variants, including Nova 2 Lite for cost-effective reasoning and everyday multimodal tasks, Nova 2 Sonic focused on speech capabilities, and related services. Nova Forge enables organizations to customize models by incorporating proprietary data early in training, while Nova Act supports AI agent technologies, including reliable browser-based automation workflows.
The portfolio aimed to give AWS customers flexible, cost-optimized options alongside third-party models available on Bedrock. Adoption reached tens of thousands of customers, though the models generally trailed the absolute leading frontier systems from specialized labs on the most demanding reasoning and multimodal benchmarks.
The Current Strategic Shift
Amazon has moved most of its higher-end and specialized Nova models – including the Premier and Omni variants, along with Reel and Canvas – into a maintenance-oriented status often referred to internally as “keep the lights on.” These models remain available and supported for existing production workloads but no longer receive major ongoing development investment.
Engineering resources and scarce high-end computing capacity are instead concentrating on Frontier Model Research. This effort, which has become a top internal priority, is developing a new flagship foundation model expected to debut at the AWS re:Invent conference in late 2026 (November 30 to December 4). The new model may still carry Nova branding. The active remaining portfolio centers on Nova 2 Sonic, Nova 2 Lite, Nova Forge, and Nova Act.
This consolidation follows organizational adjustments within Amazon’s artificial general intelligence groups. Job reductions occurred in parts of the AGI organization in July 2026, and a San Francisco-based AGI Lab – established after bringing in talent and technology from the startup Adept – was closed. Leadership of broader AGI efforts, now combined with custom silicon development and quantum computing, sits under a senior infrastructure executive. Frontier model research is led by Pieter Abbeel, who joined Amazon through the 2024 licensing and talent arrangement involving the robotics AI company Covariant.
Amazon has stated that it continually evolves its model lineup based on customer needs, continues investing in and supporting models that customers rely on, and remains committed to next-generation frontier research while providing clear migration paths.
Context Within Amazon’s Broader AI Approach
Amazon’s AI strategy has long been multi-layered rather than centered solely on proprietary frontier models. AWS Bedrock functions as a multi-vendor marketplace, offering access to models from partners including Anthropic and expanded arrangements involving OpenAI models and agent platforms. Deep partnerships provide both distribution revenue and large-scale compute commitments; Amazon has made substantial investments in Anthropic and structured multi-year compute and investment deals involving OpenAI that include significant Trainium chip capacity.
Custom silicon (Trainium series) and elevated capital expenditure – guided in the range of roughly $200 billion for 2026, heavily weighted toward AI infrastructure – underpin the infrastructure play. This positions AWS to benefit whether customers run proprietary Nova models, partner models, or third-party systems. Analyses note that Amazon’s proprietary models have not consistently matched the very highest-capability systems for the most demanding workloads. The dual approach of proprietary development plus marketplace and infrastructure depth has been described as pragmatic for enterprise customers who value optionality, cost control, security, and scalability alongside raw capability.
Industry Perspective and Implications
The move reflects a wider pattern among large technology companies of streamlining AI portfolios. Rather than sustaining parallel investments across numerous specialized model families (text, image, video, and multimodal), resources are concentrating on fewer high-priority frontier efforts capable of advancing overall capability. Compute scarcity, talent competition, and the rapid pace of progress at the leading edge make broad portfolios expensive to maintain at competitive levels.
For AWS customers, the practical impact centers on continued production support for existing Nova deployments and clear paths to newer options. Nova Forge and agent-oriented capabilities remain active development priorities, aligning with industry momentum toward reliable, enterprise-grade agents that can operate over extended workflows. The anticipated new frontier model at re:Invent will be closely watched as a potential step toward closing capability gaps while retaining Amazon’s traditional strengths in price-performance and integration.
Amazon’s recalibration does not signal reduced overall investment in AI. Capital spending on data centers, networking, and custom chips continues at elevated levels, partnerships expand access to leading external models, and Bedrock continues to broaden the available catalog. The emphasis has shifted toward disciplined prioritization: supporting what customers depend on today while directing scarce research and engineering capacity toward systems expected to define the next competitive tier. This approach leverages AWS’s distribution and infrastructure advantages even as proprietary model development becomes more focused.
In a market defined by rapid iteration and high fixed costs for frontier training runs, such concentration is a logical response. The success of the strategy will ultimately be measured by the performance and adoption of the forthcoming frontier model, the continued utility of remaining Nova tools and services, and AWS’s ability to remain the preferred platform for deploying AI at enterprise scale regardless of which underlying models customers select.
