Amazon is pulling the plug on most of its in-house artificial intelligence work. According to Bloomberg, the company is scaling back the ambitions of most of its Nova AI models and reorganizing its AI development teams to concentrate on what insiders call "highest priority" goals.
The move follows a string of disappointments. Amazon's large language models failed to win users, and its Nova family never became a household name the way ChatGPT, Claude or Gemini did. Most of its text, image and video models have now been placed in a "keep the lights on" state, meaning they are still supported but receive only the minimum resources needed to keep running.
A single frontier-model bet led by Pieter Abbeel
Instead of spreading its teams and computing power across many models, Amazon is diverting both toward one "frontier-model effort" headed by Pieter Abbeel. Abbeel directs the Berkeley Robot Learning Lab, and his robotics company Covariant was absorbed by Amazon in 2024.
At the time, Covariant's founders were working on AI models for robots, a focus that matches Amazon's wider push to automate as much of its operations as possible. The reorganization therefore points less toward a consumer chatbot and more toward robotics and automation inside Amazon's own business.
What this means for Amazon's AI strategy
The shift lands a week after Amazon closed one of its key AI offices in San Francisco, an 80-person site specializing in research on artificial general intelligence. That closure, combined with the Nova pullback, suggests the company is stepping away from the broad AI hype cycle.
Questions keep mounting over whether large language models can ever deliver a return on investment, let alone reach AGI. By freezing most of its models and backing one project, Amazon appears to be betting that a narrower, robotics-focused approach is the safer path.
