Pivot

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

co-written by eustella AI28. July 2026, 13:55

Amazon is overhauling its AI strategy. The company is winding down most of its in-house Nova models, reorganizing teams and concentrating engineers and scarce compute on a single frontier effort, according to reports by several US outlets citing people familiar with the matter. The changes follow a round of layoffs in Amazon’s Artificial General Intelligence (AGI) organization and the shutdown of AGI Lab – the research group Amazon built in 2024 after hiring most of the team behind AI startup Adept.

Nova in “keep the lights on” mode

Specifically, Amazon has begun deprecating most of its in-house flagship models: the high-end Nova Premier, the multimodal Omni model, the Reel video-generation model and the Canvas image-generation model. Internally, employees describe the status of these models as “KTLO” – short for “keep the lights on,” an engineering term for software that remains supported for existing customers but is no longer a development priority.

An Amazon spokesperson told Business Insider that the company has long supported AI models in production for extended periods because customers depend on them, and that Amazon remains committed to investing in frontier models. “AI models remain one of the most important things we’re working on, and that hasn’t changed,” the spokesperson said, adding that Amazon continually evolves its model lineup based on customer needs and provides clear migration paths.

So this is not a full exit from the brand. The remaining portfolio still includes Nova 2 Sonic and Nova 2 Lite, along with Nova Forge – a service for building and customizing models – and Nova Act, Amazon’s AI agent technology. The new flagship model could ultimately ship under the Nova label as well.

One lab fewer, one focus more

The AGI organization was created in 2023 to build foundation models and underlying technologies for Amazon products. It was led until December 2025 by longtime Alexa executive Rohit Prasad. It also housed AGI Lab, founded in December 2024 around Adept co-founder David Luan and focused on long-term research. Luan left Amazon in February; the lab was shut down as part of the latest reorganization, and the San Francisco site is being closed.

In parallel, Frontier Model Research (FMR) was built up inside the AGI organization. It is led by Pieter Abbeel, the UC Berkeley professor who joined Amazon through the acquisition of robotics startup Covariant. FMR has become the organization’s top internal priority, with resources shifting away from the existing Nova models. The new foundation model developed there is expected to debut at Amazon’s re:Invent conference in the fall.

Since December 2025, the entire AGI organization has reported to Senior Vice President Peter DeSantis, combined with the company’s custom silicon and quantum computing groups. According to the reports, DeSantis is pursuing a considerably narrower strategy than his predecessor: where Prasad ran multiple model families across text, image and video in parallel, DeSantis is concentrating talent and compute on a small number of frontier efforts. Employees report receiving little guidance about the long-term future of Nova. The layoffs surprised many, as frontier researchers had been among Amazon’s most prized technical talent; the AGI unit ran its own leveling and compensation systems to compete more aggressively for AI hires.

The other half of the equation: OpenAI on Bedrock

While Amazon pulls back on its own models, it is opening its cloud to the competition. In November 2025, AWS and OpenAI announced a multi-year strategic partnership worth $38 billion: OpenAI runs AI workloads on Amazon EC2 UltraServers with hundreds of thousands of NVIDIA GPUs (GB200 and GB300), scalable to tens of millions of CPUs. Full capacity is slated to be in place by the end of 2026, with the option to expand into 2027 and beyond.

The product layer followed in April 2026. AWS and OpenAI announced three offerings on Amazon Bedrock, initially in limited preview: OpenAI models including the frontier model GPT-5.5, the Codex coding agent, and Amazon Bedrock Managed Agents, powered by OpenAI. Customers authenticate with AWS credentials, inference runs through Bedrock, and usage can be applied toward existing AWS cloud commitments. Codex is available via CLI, desktop app and the VS Code extension; according to OpenAI, more than four million people use the coding agent every week. GPT-5.5, GPT-5.4 and Codex became generally available in June, followed by the GPT-5.6 family in July.

For Amazon, that means the models enterprise customers actually want to build with now run inside the AWS perimeter of IAM, PrivateLink, guardrails and CloudTrail – regardless of who trained them.

And Anthropic: a $100 billion cloud commitment

The entanglement with Anthropic runs deeper still. On April 20, 2026, the two companies announced an expanded partnership: Amazon is investing another $5 billion and has put up to $20 billion in milestone-based capital on the table. That brings Amazon’s cumulative Anthropic investment to roughly $13 billion. In return, Anthropic commits to spending more than $100 billion on AWS technologies over ten years and securing up to 5 gigawatts of Trainium capacity.

The foundation is Project Rainier, the jointly operated Trainium cluster that has been fully operational since October 2025. Anthropic says it uses more than one million Trainium2 chips there to train and serve Claude; close to 1 GW of combined Trainium2 and Trainium3 capacity is expected online by the end of 2026. More than 100,000 customers run Claude models on AWS, making Claude one of the most-used model families on Bedrock. Anthropic also works directly with Amazon’s chip subsidiary Annapurna Labs on the design of future Trainium generations.

What the picture adds up to

Taken together, a division of labor is emerging that Amazon has never spelled out in these terms: the company is dialing back its ambition to build the best model itself and instead maximizing its position as the infrastructure and distribution layer – through Bedrock as a model marketplace, through its own silicon, and through stakes in the labs that supply the models.

Whether that reads as a retreat or a sharpening of focus depends on the vantage point. Amazon itself talks about concentrating on what matters most to customers and points to the new foundation model announced for re:Invent. What is clear: two years ago, CEO Andy Jassy personally positioned the AGI unit as the team that would build Amazon’s most ambitious foundation models, and at last year’s re:Invent AWS presented Nova Omni 2 as its flagship multimodal reasoning model. Less than three years later, part of that line is being retired.

More clarity should come with Q2 results, which Amazon reports on July 30 – including an indication of how the roughly $200 billion in capital expenditure planned for 2026 and AI demand are tracking.