Why Is Amazon Scaling Back Its Nova AI Models?

Four Models Cut, One New Frontier Bet

Amazon is winding down active development on four of its in-house Nova AI models: Nova Premier (its most capable reasoning model), Nova Omni (multimodal), Nova Reel (video generation), and Nova Canvas (image generation). Internally, staff describe the affected models as "KTLO" — keep the lights on — meaning existing customers stay supported, but the models won't get new features or meaningful improvements going forward. Nova Premier is formally scheduled for end of life on AWS Bedrock on September 14, 2026, with Nova Canvas and two Nova Reel versions following on September 30.

Engineering talent and compute are being redirected to a new Frontier Model Research (FMR) group led by Pieter Abbeel, the UC Berkeley professor who joined Amazon through its acquisition of robotics startup Covariant. Amazon's first model from that effort is expected to debut around AWS re:Invent in late 2026 — it may or may not carry the Nova name, and no architecture, performance, or pricing details had been disclosed as of this writing.

Nova Isn't Disappearing, It's Narrowing

An Amazon spokesperson pushed back on the framing of retreat, telling reporters: "AI models remain one of the most important things we're working on, and that hasn't changed," adding that Amazon "continually evolves" its lineup based on customer needs and provides migration paths. That's a defensible position — but it doesn't change the concrete fact that four named models are no longer receiving active development.

Nova Never Became a Household Name

Amazon is concentrating scarce GPU capacity on fewer, bigger model bets rather than a broad model portfolio.

Unlike OpenAI's ChatGPT, Anthropic's Claude, or Google's Gemini, Amazon's Nova models never built strong independent brand recognition, and reportedly proved expensive to run relative to the value they delivered as cheaper open-weight rivals flooded the market through 2026. Amazon's core advantage was always cloud infrastructure rather than model research — the company remains one of Anthropic's largest backers at roughly $13 billion cumulative investment, alongside a $38 billion multi-year partnership with OpenAI for compute capacity on AWS.

If You're a Nova Customer

Who's Running Amazon's AI Strategy Now

The reorganization traces back to a December 2025 internal memo that put AWS infrastructure chief Peter DeSantis in charge of a consolidated AI group, folding in foundation-model development, custom silicon, and quantum computing under one leader. Rohit Prasad, who had led Amazon's AI efforts including the original Nova launch, departed at the end of 2025. David Luan, founder of the AGI Lab that Amazon built in 2024 after acquiring most of the team behind AI startup Adept, left in February 2026 — and that lab's 80-person San Francisco site closed entirely in July.

In their place, Pieter Abbeel — a UC Berkeley professor who joined Amazon through its acquisition of the robotics startup Covariant — now leads Frontier Model Research, the group tasked with building Amazon's next flagship model. The reshuffle followed broader layoffs across Amazon's AGI organization, reported at around 16,000 roles earlier in 2026, part of a company-wide push to concentrate resources on fewer, higher-conviction bets rather than a wide model portfolio.

Amazon's Retreat Fits a Broader Industry Shift

Amazon's pullback from a broad Nova model lineup toward a narrower, higher-conviction bet isn't happening in isolation — it reflects a pattern showing up across the AI industry through 2025 and 2026, as the initial land-grab phase of releasing a model in every size and category has given way to harder questions about which bets actually justify their compute cost. Building and maintaining a full family of models across every size tier is expensive to sustain indefinitely, and a company without a model that's genuinely winning market share in a specific category faces a real choice: keep funding a broad portfolio that isn't breaking through, or consolidate resources behind fewer models with a clearer path to competing at the frontier.

What makes Amazon's situation distinct from a pure model-strategy story is its position as a major AI infrastructure provider through AWS. Even if Nova itself never becomes a top consumer-facing model, Amazon's dominant cloud position gives it a different kind of leverage than a pure model lab has — enterprise customers already running workloads on AWS represent a large built-in audience for whatever Amazon ships next, regardless of how Nova's public brand recognition compares to ChatGPT or Gemini. That distribution advantage is likely part of why Amazon is willing to make a more concentrated bet on a next-generation frontier model rather than simply exiting the space.

Frequently Asked Questions