Amazon is pulling the plug on most of its flagship Nova AI models — and the real question is whether the company is retreating or repositioning for something bigger. The move to wind down Nova Premier, Omni, Reel, and Canvas signals a sharp strategic reset, one that trades breadth for a single, concentrated bet on a frontier foundation model.
Summary
Key takeaways
- Amazon is deprecating most flagship Nova AI models including Premier, Omni, Reel, and Canvas, putting them in “keep the lights on” mode for existing customers only.
- Resources are shifting to Frontier Model Research (FMR), a new internal effort led by Pieter Abbeel, aimed at building one competitive foundation model.
- A new flagship AI model is expected to debut at Amazon’s re:Invent conference this autumn, and could still carry the Nova name.
- Amazon closed its AGI Lab and reduced staff; Peter DeSantis has led the consolidated AI group since December, narrowing focus to fewer frontier bets.
- AWS holds compute commitments worth $138bn from OpenAI and more than $100bn from Anthropic, underlining where Amazon’s real AI leverage sits.
Amazon Winds Down Most Nova AI Models
The scope of the pullback is wider than it first appeared. According to reporting by Business Insider, Amazon has begun deprecating the high-end Premier and Omni models, the Reel video generator, and the Canvas image generator. Internally, employees have described these products as operating in “KTLO” mode — engineering shorthand for “keep the lights on” — meaning they remain supported for existing customers but no longer receive active development investment.
That framing matters. These are not products being deleted overnight. But they are effectively frozen, while the engineers and compute resources behind them move elsewhere.
What Gets Cut — and What Stays
Not everything in the Nova line is going away. Amazon confirmed it is retaining Nova 2 Lite and Nova 2 Sonic as foundation models, alongside Nova Forge, a service for building and customizing models, and Nova Act, its AI agent tool. The new frontier model that FMR is developing could eventually debut under the Nova brand as well.
Just a year ago, AWS used its re:Invent stage to unveil Nova Omni 2 as its flagship multimodal reasoning model. Less than three years after CEO Andy Jassy personally championed AGI as the team that would build Amazon’s most ambitious foundation models — prompting the creation of six new research groups — the organization is now retiring significant chunks of that same lineup.
New Focus on Frontier Model Research Led by Pieter Abbeel
The centerpiece of Amazon’s new direction is Frontier Model Research, or FMR — an internal initiative that has become the company’s top AI priority this year. Resources that once fed multiple model families are now flowing here.
One Model to Win With
FMR is led by Pieter Abbeel, who joined Amazon through its 2024 acquisition of AI robotics startup Covariant. Under his direction, the team is building a single flagship foundation model expected to debut at Amazon’s re:Invent conference this autumn, according to Reuters and Business Insider.
The logic is straightforward, if uncomfortable to admit publicly. Under former AI chief Rohit Prasad, Amazon spread its engineering resources across text, image, and video model families simultaneously. That approach produced a wide portfolio but nothing that truly competed at the frontier. Peter DeSantis, who took over the consolidated AI group in December after Prasad’s departure at the end of 2025, has since concentrated talent and scarce compute on fewer, higher-stakes efforts.
Leadership Transition and the Cost of Trying to Do Everything
Prasad’s exit was followed in February by the departure of David Luan, the Adept cofounder who led Amazon’s AGI Lab. The lab itself was subsequently shut down. AGI Lab had been established in 2024 after Amazon hired most of the team behind AI startup Adept and licensed its technology — a significant investment that is now being written off organizationally.
The pattern here reflects a broader industry tension. Building competitive frontier models requires not just talent and capital, but sustained, focused compute allocation. Spreading that across six or seven parallel model families dilutes all three. DeSantis appears to have made the unsentimental call that Amazon needed to pick a lane.
Organizational Changes and What They Signal
The AGI organization Amazon created in 2023 had also developed its own internal culture — separate leveling and compensation systems designed to compete aggressively for AI talent. Employees told Business Insider that the recent layoffs surprised many, partly because frontier model researchers had long been considered among Amazon’s most valued technical staff. The restructuring collapsed AGI under DeSantis’s broader group, which also covers silicon development and quantum computing.
Amazon’s San Francisco AGI site, an 80-person research group, has also closed, according to GeekWire.
Amazon’s Infrastructure Advantage — and the Honest Reckoning
An Amazon spokesperson pushed back on the narrative of retreat. “AI models remain one of the most important things we’re working on, and that hasn’t changed,” the spokesperson told Business Insider, adding that the company “continually evolve[s]” its model lineup based on what customers need and provides clear migration paths as models advance.
That official position is defensible, but the strategic context tells a more nuanced story. Amazon never established Nova as a widely recognized AI brand the way OpenAI, Anthropic, or Google did with their models. Its systems also proved expensive to run relative to the value they returned, at a moment when cheaper rivals were flooding the market. The competitive pressure was real.
What Amazon does have, and what no amount of model deprecation can diminish, is its infrastructure position. AWS carries compute commitments worth $138 billion from OpenAI and more than $100 billion from Anthropic. Jeff Bezos has publicly called the company’s custom Trainium chips a fourth company pillar — a multibillion-dollar business built on the premise that Amazon wins by being the engine room for everyone else’s AI ambitions.
That is a powerful position. But it also creates a strategic tension that FMR will have to resolve: Amazon is the landlord for two of its most formidable AI competitors. Whether it can simultaneously host Claude and GPT-4 on AWS while producing a frontier model that developers actively prefer over either is the bet that re:Invent will put to the test.
FAQ
Which Amazon Nova AI models are being discontinued?
Amazon is winding down most flagship Nova AI models including Premier, Omni, Reel, and Canvas. These are being placed in “keep the lights on” mode, meaning they remain supported for existing customers but are no longer an active development priority.
What is Amazon’s new focus in AI model development?
Amazon is concentrating resources on Frontier Model Research (FMR), led by Pieter Abbeel, with the goal of developing a single competitive foundation AI model rather than maintaining multiple parallel model families.
When will Amazon’s new foundation model debut?
The new flagship AI model is expected to debut at Amazon’s re:Invent conference this autumn. It could still carry the Nova brand name.
Has Amazon stopped investing in AI models altogether?
No. An Amazon spokesperson stated that AI models remain a top priority, and the company says it continually evolves its lineup based on customer needs while providing clear migration paths as models advance.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
