The company is quietly winding down its flagship internal models and pouring what is left into one big swing due later this year. Here is what is changing, and why it matters even for the AI shopping world that brands and buyers are stepping into.
Amazon is redrawing its artificial intelligence strategy, and the new plan involves doing less, not more. According to a report from Business Insider on July 28, later confirmed in the broad strokes by Reuters, the company has begun winding down most of its flagship Nova models and redirecting engineers and computing power toward a single new frontier model. It is a striking reversal for a company that, barely eighteen months ago, launched the Nova family with fanfare as an affordable alternative to the models coming out of OpenAI and its rivals.
Which models are going quiet
The list of models being pulled back is longer than it first appeared. Amazon has started deprecating Nova Premier, its highest capability language model, along with Omni, its multimodal system, Reel, a video generation model, and Canvas, an image generation tool. Inside the company, employees have described these products as being in keep the lights on mode, engineering shorthand for a product that will still be supported for existing customers but no longer receives real development effort. They are not being switched off overnight, but they are effectively frozen while the talent behind them moves elsewhere.
For the developers who build on these models through Amazon’s Bedrock service, the wind down comes with firm dates. Amazon has classified several of the affected models as legacy, with Nova Premier scheduled to reach end of life on September 14 and both Nova Canvas and two versions of Nova Reel set to follow at the end of September. Omni sits in a slightly different spot, with no published shutdown date yet, which means its reduced status has not hardened into a confirmed cutoff. Legacy models remain available to current customers for now, but new customers cannot pick them up.
Not everything in the lineup is disappearing. Amazon has confirmed it will keep investing in a narrower set of models, including Nova 2 Lite and the speech focused Nova 2 Sonic, alongside Nova Forge, a service that lets customers build and customize their own models, and Nova Act, an agent tool built to operate browsers and software interfaces. Those survivors are not drop in replacements for everything being retired, but they signal where Amazon still sees value.
One big bet instead of many small ones
The resources freed up by the wind down are flowing into a group called Frontier Model Research, led by Pieter Abbeel, a University of California, Berkeley professor who came to Amazon through its acquisition of the robotics startup Covariant. That team has become the company’s top AI priority, and it is building a new flagship foundation model that is expected to debut at Amazon’s re:Invent conference later this year. In a twist that captures how fluid the whole effort is, the new model could yet launch under the same Nova brand Amazon is currently thinning out.
The strategy shift did not appear out of nowhere. It follows the closure of Amazon’s AGI Lab in San Francisco after one of its cofounders departed earlier in the year, along with a round of layoffs among frontier researchers who had been among the company’s most prized technical hires. Late last year, Amazon folded its broader AGI organization under a senior cloud executive, Peter DeSantis, who also oversees the company’s custom chip and quantum computing work. Seen in that light, deprecating the Nova models is the product level proof of a retreat that until now had shown up mainly as shuttered offices and job cuts.
Amazon, for its part, has framed the changes as routine portfolio management. A company spokesperson told Reuters that it continually evolves its lineup based on what customers need, and stressed that it is still supporting the models customers rely on today while investing in the next generation of frontier research. That is the optimistic reading. The blunter reading, offered by much of the coverage, is that Amazon has concluded it cannot win by spreading its effort thin across many specialized models, and is consolidating around one serious attempt to build something genuinely competitive.
The awkward position of hosting your rivals
What makes Amazon’s bet unusual is the position it occupies in the wider AI market. Even as its own models stumbled, Amazon has built an enormous business as the infrastructure provider for everyone else’s AI ambitions. AWS carries massive computing commitments from the very companies Amazon is competing with, including tens of billions of dollars in arrangements with OpenAI and a deep, multibillion dollar relationship with Anthropic. Amazon’s investment in Anthropic was large enough that a gain on its value added more than $50 billion to the company’s reported profit in the most recent quarter.
That creates a genuine tension. Amazon is, in effect, the landlord for two of its most formidable competitors, hosting their models on AWS while trying to ship a frontier model that developers would actively choose over either. Its custom Trainium chips, which Jeff Bezos has called a fourth pillar of the company, are central to that pitch, the idea being that Amazon wins by being the engine room for the entire industry. Whether it can be both the neutral platform and the ambitious challenger at the same time is the question the new model will have to answer.
Why sellers and shoppers should care
On the surface this looks like a cloud and developer story, far removed from the day to day of selling products online. But the reshuffle is happening directly beneath the AI features that are reshaping how people shop. Amazon has leaned hard into AI on the storefront, from its Rufus shopping assistant to automated tools that generate listing content and creative for brands, and agent technology like Nova Act hints at a near future where software does more of the browsing and buying on a shopper’s behalf. The plumbing under those features is what is changing here. For the brands that spend their days optimizing for Amazon’s AI surfaces, the lesson is less that those surfaces are going away and more that they will keep shifting underneath, which puts a premium on staying adaptable rather than betting on any one tool.
For a company of Amazon’s size, admitting that an expensive strategy is not working and starting over is not a small thing. The Nova wind down is that admission, made in the quiet language of end of life dates and reassigned engineers. Amazon is trading breadth for focus and leaning on the one advantage no model shutdown can touch, its infrastructure. Later this year at re:Invent, the market will get its first real look at whether the single big swing was worth everything the company gave up to take it.
Reporting sources: Business Insider, Reuters, eWeek, Neowin, The Next Web, MLQ, The Spokesman Review, and Amazon.
