Amazon Scales Back Its Own AI Models to Focus on Infra, OpenAI and Anthropic

Amazon just admitted, without quite saying it out loud, that building the best AI model itself isn't the plan anymore. Multiple US outlets report the company is deprecating most of its flagship Nova lineup, including the high-end Nova Premier, the multimodal Omni, and the Reel and Canvas generation models. Internally staff call this status KTLO, engineer-speak for keep the lights on: still supported for existing customers, no longer where the energy goes. The move follows layoffs inside Amazon's AGI organization and the shutdown of AGI Lab, the research group built in late 2024 around most of the team Amazon poached from AI startup Adept.

The reorganization tells a story about priorities. AGI Lab's David Luan left Amazon in February, and its San Francisco office is closing for good. Meanwhile Frontier Model Research, led by UC Berkeley's Pieter Abbeel, has become the internal priority, with a single new foundation model expected to debut at re:Invent this fall. Since December, the whole AGI operation reports to Peter DeSantis, who also oversees Amazon's custom silicon and quantum computing groups, and who is running a noticeably narrower playbook than his predecessor Rohit Prasad. Where Prasad spread bets across text, image and video models simultaneously, DeSantis is pooling talent and scarce compute into one shot rather than several.

While Amazon trims its own model ambitions, it's rolling out the welcome mat for rivals. AWS and OpenAI struck a $38 billion, multi-year compute deal in November, putting OpenAI workloads on EC2 UltraServers packed with hundreds of thousands of Nvidia GB200 and GB300 GPUs. By April, that turned into product: GPT-5.5 and the Codex coding agent, used by more than four million people weekly, landed on Amazon Bedrock alongside a new managed-agents offering. Customers log in with AWS credentials, run inference through Bedrock, and the spend counts toward their existing AWS commitments — meaning OpenAI's models now live inside Amazon's security perimeter no matter who trained them.

The Anthropic relationship goes even further. On April 20, Amazon added another $5 billion in direct investment plus up to $20 billion in milestone-based capital, pushing its total stake in Anthropic to roughly $13 billion. Anthropic, in turn, pledged more than $100 billion in AWS spending over a decade and locked in up to 5 gigawatts of Trainium capacity. Project Rainier, the shared Trainium cluster that went fully live in October, already runs over a million Trainium2 chips training and serving Claude, with close to 1 gigawatt of combined Trainium2/3 capacity due online by the end of 2026. More than 100,000 customers already run Claude on AWS, making it one of Bedrock's most-used model families.

Add it up and a division of labor emerges that Amazon has never quite said out loud: let OpenAI and Anthropic fight over frontier bragging rights, and make sure every dollar of that fight flows through AWS silicon, AWS security tooling and AWS billing. Two years ago Andy Jassy framed the AGI unit as the team building Amazon's most ambitious models. Less than three years later, a chunk of that ambition is being quietly retired, and Amazon's Q2 earnings on July 30 should show whether the roughly $200 billion in planned 2026 capex is chasing that new strategy or just paying for someone else's models to run in Amazon's basement.