Amazon AI Strategy Overhaul: Nova Models Dropped for Single Frontier Bet
Amazon is executing a sweeping AI strategy overhaul, pulling multiple flagship models from its Nova lineup including the high-end Premier and Omni tiers, the Reel video generator, and the Canvas image tool, to redirect engineering resources and compute budget toward a single next-generation foundation model. The affected models remain accessible to existing customers in a maintenance-only mode, but they will no longer receive active feature development.
The restructuring, first reported by Business Insider on July 28 and confirmed by Reuters, is the most significant shift in Amazon's AI strategy since the Nova brand launched less than two years ago. Behind the cuts are layoffs within the company's Artificial General Intelligence organization this month, the closure of AGI Lab, a research unit established in 2024 from the Adept acquisition, and a broader consolidation of AI, custom silicon, and quantum computing units under senior vice president Peter DeSantis.
The Frontier Model Bet Behind Amazon's AI Strategy Overhaul
Resources are now flowing to Frontier Model Research, or FMR, a group led by Pieter Abbeel. Abbeel joined Amazon through the Covariant acquisition and has been given primary responsibility for developing what the company intends as its new flagship foundation model. According to internal plans, that model is slated for a debut at Amazon's re:Invent conference in the fall of 2026.
FMR replaces a multi-model strategy that had Amazon maintaining separate foundation models for text reasoning with Nova Premier and Nova Omni, video generation with Reel, and image generation with Canvas. Internal teams described the deprecated models as existing in a keep-the-lights-on state, with engineering attention shifted entirely to the frontier effort.
Not all Nova models are being wound down. Amazon continues to support Nova 2 Sonic, Nova 2 Lite, Nova Forge, and Nova Act, suggesting the company is retaining a narrow application layer even as it consolidates its core foundation model development. The Amazon AI strategy overhaul preserved these products because they serve more targeted enterprise use cases, unlike the broad frontier models being retired.
The Economics of Consolidation
The pivot reflects a calculation that even Amazon Web Services, a business that generated over $100 billion in annual revenue, cannot sustainably fund parallel model families across text, image, and video modalities at frontier quality. Training and operating top-tier models across multiple domains requires enormous GPU clusters, data pipeline investment, and research talent. By concentrating those resources on a single effort, Amazon is following a path similar to what other large AI players have done, though the cutbacks are more abrupt than the gradual transitions seen at some competitors.
The departure of former AGI head Rohit Prasad in December 2025 created the opening for DeSantis to merge what had been separate AI, silicon, and quantum computing divisions. The organizational change narrowed strategic focus and placed the full weight of Amazon's custom chip efforts, including its Trainium and Inferentia lines, behind the frontier model push.
What Survives and What Changes
For enterprise customers who have built workflows on Nova Premier, Nova Omni, Reel, or Canvas, the immediate change is limited. Amazon has committed to maintaining service availability for existing customers, and the models will continue to serve inference requests. But the absence of new features means those products will eventually lag behind alternatives that are still under active development.
The retained models, Nova 2 Sonic and Nova 2 Lite for text, Nova Forge for enterprise customization, and Nova Act for agentic tasks, provide a bridge for current users. Whether Amazon expands this smaller portfolio after the frontier model launches at re:Invent, or strips it down further, is an open question.
Why this matters
Amazon's decision to wind down most of its Nova portfolio signals that the cost of maintaining breadth in AI model development has become prohibitive even for the largest technology companies. For enterprise buyers, this means evaluating AI platform choices now involves not just capability but longevity. A model family that seems comprehensive today can be pared back rapidly when the economics shift. The fall re:Invent conference will show whether Amazon's single-bet strategy can produce a model that competes with offerings from Google, OpenAI, and Anthropic, or whether the consolidation has left it playing catch-up in every modality at once.
Related Articles
- AWS Cuts AI Costs with Nova Model Distillation
- Amazon Bedrock Integrates OpenAI GPT OSS and NVIDIA Nemotron to Diversify Enterprise AI Options
- Beyond Bigger Models: The Sovereign Efficiency Pivot Reshaping AI's Future
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team.