Amazon quietly stops developing most of its own AI models
Nova Premier, Omni, the Reel video model and the Canvas image generator go into keep-the-lights-on mode. Resources shift to a new Frontier Model Research group under Pieter Abbeel.
Amazon is scaling back most of its in-house Nova models, Business Insider reports, citing people inside the company. The cuts hit the flagship Nova Premier and Omni models, the Reel video model and the Canvas image generator. The models stay online in what insiders describe as “keep the lights on” mode for existing customers, meaning they still run but nobody is making them better.
The pullback follows layoffs in Amazon’s AGI division and the closure of the AGI Lab the company founded in 2024 after acquiring the startup Adept. Amazon first unveiled the Nova family in December 2024, positioning it as a cost-and-performance middle ground rather than a frontier contender. Eighteen months later, that middle ground turned out to be the worst place to stand: too expensive to build, not good enough to win on capability, and not cheap enough to win on price against a flood of open-weight models.
Amazon is not leaving the field. Resources now flow to a new Frontier Model Research group led by Pieter Abbeel, the Berkeley robotics researcher who joined through Amazon’s acquisition of Covariant. A new foundation model is expected at the re:Invent conference this autumn. A company spokesperson told Business Insider that AI models remain one of Amazon’s most important projects.
There is a second strategy running alongside the first, and it may be the real one. Amazon is among the largest backers of both Anthropic and OpenAI, with investments that could exceed $100 billion combined. If you own a large cloud business and significant stakes in two of the strongest model labs, the case for also building your own mid-tier models is thin. You make money when customers run inference on AWS, and it matters less whose model they run. Building a fourth-best model to compete with your own investees is an expensive way to have a strategy.
That said, “we invest instead of build” has a dependency problem baked in, which is presumably why the Frontier group exists at all. Amazon appears to be trading breadth for one serious attempt.
What this means for you: If you use Nova models through AWS Bedrock, they keep working, but plan on migrating eventually, because a model in maintenance mode falls behind the field within months. If you follow the industry, this is the clearest sign yet that the middle of the model market is closing. Frontier labs take the top, open-weight models take the cheap end, and the merely-competent proprietary model in between has nowhere left to sell. Expect more of these announcements, phrased just as gently.
Sources
Source: https://www.businessinsider.com/amazon-ai-strategy-nova-models-frontier-model-research-2026-7
Google DeepMind has broken up the AlphaFold team, and a quarter of its authors have left
The team behind the Nobel-winning protein model was reassigned over the past year. Nobel laureate John Jumper and two colleagues are heading to Anthropic.
