Amazon launched its Nova family of AI models at re:Invent in December 2024 with much fanfare: Nova Premier, Nova Omni, Nova Reel, Nova Canvas.
The pitch was that Amazon was building affordable, capable alternatives to the models from OpenAI, Anthropic, and Google. It was a broad lineup designed to show the market that AWS wasn’t just cloud infrastructure anymore. It was an AI company.
Less than two years later, Amazon is quietly pulling most of those models. Reuters reported on July 28 that Amazon has begun deprecating most of its flagship Nova models and reorganizing its AI teams around a single frontier-model effort.
The overhaul is the most significant reset of Amazon’s AI strategy since the Nova launch.
Which Amazon Nova AI models are being discontinued and what survives
The list of what’s going away is significant. Nova Premier and Nova Omni, Amazon’s high-end foundation models, are being moved into what engineers internally call KTLO status, short for “keep the lights on.”
That means they stay supported for existing customers but are no longer a development priority. Nova Reel, the video-generation model, and Nova Canvas, the image-generation model, are in the same situation.
More Amazon:
- Bank of America doubles down on Amazon shares after Prime Day
- Amazon’s $8.3 billion Prime Day sends Wall Street a warning
- Amazon Prime Day gives Wall Street a $22B reason to take notice
What survives is a much smaller set. Nova 2 Lite and Nova 2 Sonic remain active. Nova Forge, which lets customers build and customize their own models using Amazon’s technology, stays. Nova Act, Amazon’s AI agent platform, continues in active development, according to Reuters.
“As with any AI portfolio, we continually evolve our model lineup based on what customers need,” an Amazon spokesperson told Reuters. “We’re continuing to invest in and support the Nova models our customers rely on today. We’re also investing in the next generation of frontier model research.”
Amazon’s new AI frontier push and what it’s building
The resources being pulled from the deprecated Nova models are going to a new internal group called Frontier Model Research, or FMR. It’s led by Pieter Abbeel, a UC Berkeley professor who came to Amazon through its acquisition of AI robotics startup Covariant in 2024. FMR has become the top internal priority inside Amazon’s AGI organization this year.
The team is building a new flagship foundation model expected to debut at Amazon’s re:Invent conference later in 2026. It may still carry the Nova brand when it launches, but it represents a fundamentally different approach than the broad portfolio Amazon previously maintained.
Instead of several specialized models for text, image, audio, and video, Amazon is concentrating on a single frontier-scale system designed to compete directly with the top models from OpenAI, Anthropic, and Google.
The AGI organization itself has also been restructured. It now reports to Peter DeSantis, a longtime Amazon cloud executive who also oversees the company’s custom silicon and quantum computing efforts.
Several senior AGI executives have departed over the past year, including Rohit Prasad, who had been the public face of Amazon’s AI efforts since the Alexa era.
Chris/Getty Images
Why Amazon is cutting AI models rather than expanding them
The honest answer is that Nova wasn’t generating the attention Amazon needed it to. While OpenAI, Anthropic, and Google were dominating the conversation around frontier AI, Amazon’s models were getting used by enterprise customers through AWS but weren’t breaking through as a brand or a benchmark story.
The economics weren’t working, either. Running and developing multiple foundation model families across text, image, audio, and video is expensive, and spreading engineering talent across all of them dilutes the results, according to Neowin.
The shutdown of AGI Lab makes the same point. Amazon built that research group in 2024 after hiring most of the team from AI startup Adept. It was meant to accelerate Amazon’s AGI research. Less than two years later, it’s closed, and the people who remained have been absorbed into FMR or moved elsewhere in the organization.
Amazon has always been better at being the picks-and-shovels play in AI than at being the AI company itself. Its investment in Anthropic, its cloud infrastructure for AI workloads, and its chip development through Trainium and Inferentia have all been performing. Its own models, less so. This reorganization is an acknowledgment of that reality.
What Amazon’s AI strategy overhaul means for AWS customers, AMZN investors
For AWS customers currently using Nova Premier, Nova Omni, Nova Reel, or Nova Canvas, the KTLO designation means their workloads stay supported. They’re not being cut off tomorrow.
But they also shouldn’t expect meaningful improvements to those models going forward. If they want Amazon’s best capabilities, they’ll eventually need to migrate to whatever emerges from FMR at re:Invent.
For investors, this is actually a more disciplined story than the one Amazon was telling before. Spreading engineering talent and compute across a dozen model families is expensive and produces mediocre results across the board.
Concentrating everything on a single frontier effort is how you build something that can compete with GPT-5 and Gemini. Whether Abbeel’s team can deliver that at re:Invent is the question investors will be watching.
Amazon reports earnings on July 30. Andy Jassy will almost certainly be asked about the AI reorganization on the call. What he says about the Nova deprecations, the FMR timeline, and how Amazon thinks about competing at the frontier will matter more than almost anything else in the report.
The strategy has changed. The market will want to know if the new one is working.
Related: Amazon CEO just made things uncomfortable for Anthropic
