Amazon Rethinks AI Strategy, Deprecates Most Nova Models to Focus on New Frontier Effort

Amazon Winds Down the Nova Portfolio and Bets on a New Frontier Model

Amazon is reshuffling its artificial intelligence ambitions, deprecating most of its own Nova model lineup and concentrating engineering firepower behind a new, undisclosed flagship foundation model. According to people familiar with the matter, the company has begun winding down the high-end Nova Premier and Nova Omni text models as well as its Reel video-generation and Canvas image-generation services. Internally, these products are considered in “keep the lights on” mode—still supported for existing customers but no longer receiving major development investment.

The strategic shift follows layoffs last week inside Amazon’s Artificial General Intelligence (AGI) organization and the closure of AGI Lab, a research unit formed in 2024 after Amazon hired much of the team behind AI startup Adept. The company is now funnelling resources into a new initiative called Frontier Model Research (FMR), led by Pieter Abbeel, the renowned researcher who joined Amazon through its acquisition of robotics AI startup Covariant. FMR has become the top priority for the AGI group this year, with the goal of unveiling a new foundation model at Amazon’s annual re:Invent conference, typically held in the fall.

An Amazon spokesperson told Business Insider that supporting AI models for extended periods is standard practice because customers rely on them, and stressed that the company “remains committed to investing in frontier models.” Some parts of the Nova brand will survive: the Nova 2 Sonic and Nova 2 Lite foundation models, the Nova Forge customization service, and Nova Act agent technology are still under active development. The new FMR model could even launch under the Nova name. But the broader trend is unmistakable—a move away from the sprawling, multi-model portfolio Amazon previously championed toward a sharper, more concentrated approach.

Why Peter DeSantis's Focused Strategy Marks the End of Amazon's AI Diversification

The Quiet Death of Amazon’s Broad Model Strategy

Until now, Amazon followed a scattergun AI model strategy, building separate families for text, image, and video generation. Nova Premier and Omni were positioned as flagship reasoning models, while Reel and Canvas targeted media creation. By putting most of those on KTLO status, Amazon is signalling that maintaining such a wide array of custom foundation models was draining resources and failing to produce a clear competitive advantage. The move also reflects an internal power shift: after AGI head Rohit Prasad departed in December 2025, new SVP Peter DeSantis consolidated the AI, silicon, and quantum groups and pushed for a far leaner approach. Multiple sources say DeSantis wants to direct scarce computing capacity and top engineering talent toward just one or two frontier efforts, rather than sustaining half a dozen parallel model lines.

What Pieter Abbeel’s Mandate Really Means

Abbeel is best known for his work in robotics and reinforcement learning, not large-language models. His appointment to lead FMR suggests Amazon may be betting on a different kind of foundation model—possibly one that integrates robotic or autonomous capabilities, leveraging Covariant’s IP. That would differentiate Amazon’s offering from GPT-style text models and align with the company’s logistics and warehouse automation strengths. But it is also a high-risk gamble: for AWS customers used to plug-and-play text and image APIs via Bedrock, a radical departure could leave a gap that competitors like Microsoft and Google are eager to fill.

The Customer and Competitive Fallout

AWS has heavily promoted Nova models as a key part of its AI stack. Enterprises that adopted Nova Premier or Omni for critical workloads now face uncertainty. Amazon promises “clear guidance and migration paths,” but no timeline has been shared publicly. This raises the risk of customer defections to OpenAI’s GPT models on Azure or Google’s Gemini, especially because re:Invent is still months away. At the same time, a concentrated, well-executed frontier model could eventually give Amazon a stronger hand. The retrenchment mirrors a broader industry pattern: even well-funded players are finding that maintaining a family of cutting-edge models is extraordinarily expensive and that focus often beats diversification.

Internal Turmoil and Talent Retention

The layoffs and reorganization have rattled Amazon’s AI talent. AGI operated with its own levelling and compensation systems to compete for researchers, so that talent was highly prized. The closure of AGI Lab—built around the Adept acquisition—and departure of its leader David Luan sends a signal that Amazon’s long-term research bets are being pruned as well. Whether the company can retain the engineers needed to deliver a transformational model while going through such upheaval is an open question.

What AWS Customers and Investors Should Do as Amazon Deprecates Nova Models

• AWS customers using Nova Premier, Omni, Reel, or Canvas should immediately review their dependency. Amazon has confirmed it will support existing workloads, but migration planning needs to begin now. Check the roadmap for Nova 2 Lite and Sonic as potential interim replacements, or consider re-architecting applications around third-party models available via Bedrock.
• Watch for model end-of-life announcements and migration tooling. Amazon’s spokesperson promised “clear guidance and migration paths.” Expect these details to emerge in the lead-up to re:Invent. Document your usage and set internal deadlines before pricing, latency, or feature gaps force a rushed switch.
• For investors, the re:Invent conference—likely late November 2026—is the pivotal event. The capabilities, pricing, and customer reception of the new frontier model will determine whether this strategic pivot restores AWS’s AI momentum or opens the door further for Microsoft and Google. Near-term, signs of customer churn in AWS’s AI revenue lines should be monitored.
• Competitors should note the window of opportunity. Amazon’s retrenchment leaves a gap in the market for enterprise-grade foundation models for at least one or two quarters. That said, a successful launch from Abbeel’s team could reset competitive dynamics, so the risk is symmetrical.