Amazon Nova is getting gutted. Amazon is winding down most of its flagship in-house models, closing the San Francisco lab that was supposed to build its AGI future, and betting the whole thing on one frontier effort plus the cloud.
Business Insider reported the model cuts on July 28. The official AWS Bedrock lifecycle page documents the end-of-life process for models.
According to the AWS Bedrock lifecycle page, Nova Premier is set for end-of-life in September, and Nova Canvas and Nova Reel will follow on September 30.
This is the clearest signal yet that Amazon quit the model arms race it never really led. And if you build on Bedrock, you have a migration clock running.
The Amazon Nova models being cut
The casualties are the ones that mattered. Nova Premier, the high-end multimodal model with a 1M token context window, is done. So are Nova Omni, the multimodal workhorse; Nova Reel, the video generator; and Nova Canvas, the image generator.
Employees reportedly call the status KTLO, short for keep the lights on. The models still run for existing customers, but they get no real development.
What survives appears to be a much thinner lineup. Business Insider reported that the remaining portfolio includes Nova 2 Sonic and Nova 2 Lite, Nova Forge for custom model building, and Nova Act for AI agents. That’s the cheap tier plus the tools that plug into the AWS machine, not a full flagship family.
The AGI Lab is gone
Amazon shut its San Francisco AGI Lab as part of the cuts. GeekWire confirmed the site closure, which The Information first reported.
The leadership exits tell the story. David Luan, the Adept co-founder who ran the lab, left in February. Rohit Prasad, the executive who oversaw AGI, left at the end of last year.
Peter DeSantis now runs a combined AI, silicon, and quantum org. CNBC reported the AGI layoffs on July 22.
One frontier bet, not many
The remaining engineering firepower is consolidating under a new frontier-model effort, known internally as Frontier Model Research (FMR), with Business Insider reporting that it is led by Pieter Abbeel, the Berkeley roboticist who joined Amazon through the Covariant acquisition.
A new Amazon Nova flagship is reportedly expected at re:Invent later this year, and it might still carry the Nova name.
That’s the tell. Amazon isn’t quitting models entirely. It’s betting that one focused effort beats six half-funded ones.
Jassy’s original AGI push, launched in 2023, spread resources across text, image, and video families. DeSantis is collapsing that into a single priority.
The infrastructure paradox
Here’s the part that makes this funny. Amazon is simultaneously cutting most of its own models and raising its AI infrastructure spending to unprecedented levels. AWS revenue is surging, with cloud growth hitting its fastest pace in years.
Amazon has instead invested heavily in cloud infrastructure and partnerships, such as its significant investment in Anthropic, to offer models like Claude on Bedrock. The Nova cuts underscore that focus.
AWS even announced a $1 billion Forward Deployed Engineering group in June, embedding engineers with customers like the NBA and Southwest Airlines to build agent systems. That’s where the money and the headcount are going. Not the leaderboard.
What this means for Bedrock users
If you’re running on Nova Premier, Reel, or Canvas in production, the deadline is real. AWS explicitly says migration won’t happen automatically, so you’ll need to plan your own move. With the end-of-life dates looming, you have barely a month and a half to migrate. If you haven’t started a migration plan, start now. The remaining Nova lineup handles a lot of workloads, and Claude and other models on Bedrock are the obvious fallback for reasoning-heavy jobs. But nobody should be writing new code against a model with a September EOL.
Bottom line
This is the most honest move Amazon has made in AI. It never won a model battle, and it just admitted it.
The bet now is that the customer machine matters more than the benchmark, and Amazon’s massive infrastructure spending says it believes that bet pays off.
Microsoft is openly competing with OpenAI and Anthropic with its own MAI models, and Google and Meta keep shipping flagships. Amazon chose a different lane: sell the rails, let everyone else fight over the trains.
Whether Abbeel’s team delivers something credible at re:Invent is the open question. If it does, Amazon gets a cleaner story. If it doesn’t, this retreat starts to look less like focus and more like exit. Either way, the message to builders is already clear: don’t build your roadmap on Amazon Nova.
