Amazon is winding down most of its Nova AI models to bet on one frontier model

Amazon is winding down most of its flagship Nova AI models. It is betting what remains on a single, bigger effort to build one genuinely competitive foundation model. The move gives shape to a retreat that until now looked mostly like closed offices and job cuts.

The company has begun deprecating most of its in-house Nova line, Business Insider first reported. That includes the high-end Premier and Omni models, the Reel video generator and the Canvas image generator. Some staff call this “KTLO” mode, short for “keep the lights on.” The models stay supported for existing customers, but are no longer a development priority.

This is the concrete update to a story we first covered months ago. Amazon shut its AGI Lab and quit a race it never led. Back then, the news was a closed lab and layoffs. Now it is clear which products are being sacrificed, and what they are being sacrificed for.

One big bet instead of many

Resources are shifting to a new effort called Frontier Model Research, or FMR, led by Pieter Abbeel. He joined Amazon through its 2024 acquisition of the robotics startup Covariant. A new flagship model is expected to debut at Amazon’s re:Invent conference this autumn, Reuters reported. It could still carry the Nova name.

The logic is focus. Under previous AI chief Rohit Prasad, Amazon spread itself across text, image and video models. Peter DeSantis, who took over the consolidated group in December, has concentrated talent and scarce compute on fewer frontier bets. Prasad left at the end of 2025, and AGI Lab founder David Luan departed in February.

Nova is not vanishing. Amazon says several products will stay. It is keeping Nova 2 Lite, Nova 2 Sonic, the Nova Forge customisation service and the Nova Act agent tool. The company’s San Francisco AGI site, an 80-person research group, has closed, GeekWire reported.

Where Amazon actually wins

Amazon never made Nova a household name the way OpenAI, Anthropic and Google did with their models. Its systems also proved expensive to run for the value they returned. Cheaper rivals piled on the pressure, part of a wider shift to cut-price models.

So Amazon leaned into what it does best: infrastructure. AWS is the landlord for much of the industry, with compute commitments worth $138bn from OpenAI and more than $100bn from Anthropic. Its custom Trainium chips are now a multibillion-dollar business that Jeff Bezos calls a fourth company pillar.

An Amazon spokesperson rejected the idea of a retreat. “AI models remain one of the most important things we’re working on, and that hasn’t changed,” the spokesperson said, adding that the company “continually evolve[s]” its lineup around what customers need.

The harder question is the one Nova never answered. Amazon can host everyone else’s models and sell the chips underneath them. Whether it can also build a frontier model that developers actively choose over Claude, Gemini or GPT is what re:Invent will test.