Amazon Is Winding Down Most of Its Nova AI Models. Stop Betting Your Creative Workflow on Them.

Business Insider reported in late July that Amazon is winding down most of its flagship Nova AI models, including the high-end Premier and Omni models, the Reel video generator, and the Canvas image generator. Internally, employees describe those models as being on KTLO status, engineer shorthand for “keep the lights on,” meaning they stay supported for existing users but get no further development. Amazon is redirecting its research firepower toward a single frontier model and is expected to show a new flagship at re:Invent this autumn, possibly still under the Nova name.

Most of the coverage framed this as an AI horse-race story about whether Amazon can compete with OpenAI and Anthropic. That is not your problem. Your problem is that two of the models on the maintenance list, Nova Canvas and Nova Reel, are the exact tools Amazon has been telling sellers to use to generate ad images and video inside the Ads console. And Rufus, the shopping assistant your listings are increasingly discovered through, runs on the same in-house model family that is being consolidated.

The recommendation

Treat Amazon’s free generative creative tools as frozen, not as a roadmap. Keep producing and owning your core image and video assets outside Amazon, and do not build an advertising workflow that depends on Nova Canvas or Nova Reel getting better. On the discovery side, stop over-tuning to how today’s Rufus behaves and put the effort into durable, machine-readable listing fundamentals, because the model underneath Rufus is going to be replaced. This is a moment to de-risk your dependence on Amazon’s AI stack, not to lean harder into it.

What Amazon actually did

Amazon launched Nova at re:Invent in late 2024 as its own family of foundation models on Bedrock. Canvas handled image generation, Reel handled video, and the larger Premier and Omni models handled reasoning and multimodal tasks. Over the following year Amazon pushed these into seller-facing surfaces, most visibly in Amazon Ads, where Canvas and Reel let sellers spin up display and video creative, run video advertising for the first time, and generate creative down to the keyword level without a production budget.

The reported change does not delete those tools overnight. KTLO means they keep running for now. What it removes is the assumption that they are on an upward trajectory. A tool in maintenance mode does not get better quality, new formats, or fixes to the quirks you are working around today. It is a static utility, and it can be sunset later with little notice once the new frontier model lands.

Why this hits sellers specifically

Two exposures matter here. The first is creative. If your ad production leans on Amazon’s built-in image and video generation because it is free and fast, you have quietly outsourced a core asset to a tool Amazon just deprioritized. The generated output was already generic enough to blend in with every other seller using the same model. Now it will not improve either. Brands that own a real creative library, produced independently and reused across Amazon, Walmart, and their own site, are insulated. Brands that let Amazon generate their creative on demand are exposed to whatever Amazon decides to do with these models next.

The second exposure is discovery. Amazon has spent two years telling sellers to optimize for AI shopping, first through Rufus and now through the Alexa for Shopping surfaces. That AI layer runs on the model family Amazon is now rebuilding. Any tactic finely tuned to how the current model reads and ranks your listing is built on sand, because a new frontier model will read and rank differently. The winners will not be the sellers who reverse-engineered this quarter’s Rufus. They will be the sellers whose product data is clean enough that any model, current or next, can parse it correctly.

What to do now

Audit where Amazon’s generative tools sit in your workflow and pull your creative production back in house or to a partner. Keep your hero images, lifestyle shots, and video as owned assets you control and can deploy anywhere, so no single platform’s model decisions can degrade your storefront.

On listings, put the work into structure rather than model-chasing. Accurate, complete attributes, clean titles within the new 75-character limit, the Item Highlights field filled with the specs that used to live in long titles, and backend search terms that describe the product in plain language. This is the input every AI shopping surface depends on, and it is the one part of the system that does not reset when the model changes.

Then hold your advertising strategy on fundamentals that do not care which model is live: tight campaign structure, disciplined bids, and creative that actually converts. If you have been using Nova-generated creative as a crutch, this is the push to fix it before peak season, not after.

The takeaway

Amazon consolidating its model lineup is a rational engineering decision, and the new frontier model may well be better than what it replaces. But every platform shift like this transfers risk to the sellers who wired their operation to the old tools. The lesson is the same one that applies every time Amazon changes a program: own your assets, keep your data clean, and never let a free platform tool become a load-bearing part of your business. The sellers who get hurt by the Nova wind-down are the ones who let Amazon generate their creative and optimized their listings for a model that is already on its way out.

If you want a read on where your creative and listing setup is over-exposed to Amazon’s AI tooling, and what to shore up before re:Invent and Q4, book a strategy assessment and we will map it against your catalog.