Amazon is winding down most of its flagship Nova AI models, consolidating its artificial intelligence ambitions into a single, larger bet on a genuinely competitive foundation model. The move gives concrete shape to a retreat that until recently looked mostly like closed offices and job cuts. According to Business Insider, the company has begun deprecating most of its in-house Nova line, including the high-end Premier and Omni models, the Reel video generator, and the Canvas image generator. Some staff internally refer to this as “KTLO” mode, short for “keep the lights on.” The models will remain supported for existing customers, but they are no longer a development priority.
This is the latest update to a story that first emerged months ago. Amazon shut its AGI Lab and essentially withdrew from the race it never led. At that time, the news was a closed lab and layoffs. Now the picture is clearer: specific products are being sacrificed, and the resources are being redirected toward a focused effort to build a single frontier model that can compete with the best in the industry.
One big bet instead of many
Resources are shifting to a new initiative called Frontier Model Research, or FMR, led by Pieter Abbeel. Abbeel joined Amazon through its 2024 acquisition of the robotics startup Covariant, where he was a co-founder and a prominent figure in the AI and robotics research community. Under this new initiative, a flagship model is expected to debut at Amazon’s re:Invent conference this autumn, as reported by Reuters. That model could still carry the Nova name.
The logic behind this consolidation is focus. Under previous AI chief Rohit Prasad, Amazon spread its efforts across text, image, and video models, aiming to cover a broad spectrum of generative AI use cases. However, this approach proved costly and diffused the company’s talent and computing resources. Peter DeSantis, who took over the consolidated group in December, has concentrated talent and scarce compute on fewer frontier bets. Prasad left at the end of 2025, and AGI Lab founder David Luan departed in February. These leadership changes signal a strategic pivot away from the earlier, more expansive approach toward a more disciplined, results-oriented strategy.
Nova is not vanishing entirely. Amazon has said that several products will stay. The company is keeping Nova 2 Lite, Nova 2 Sonic, the Nova Forge customisation service, and the Nova Act agent tool. These are perhaps the more commercially viable or technically differentiated offerings. The company’s San Francisco AGI site, an 80-person research group, has also been closed, according to GeekWire. This closure is a tangible sign of the downsizing that has accompanied the strategic shift.
Where Amazon actually wins
Despite its investments, Amazon never made Nova a household name the way OpenAI, Anthropic, and Google did with their models. The company’s AI models were not widely adopted by developers or enterprises seeking state-of-the-art performance. Moreover, its systems proved expensive to run for the value they returned. Cheaper rivals piled on the pressure, part of a wider industry shift toward low-cost, highly efficient models. This economic reality made it difficult for Amazon to justify maintaining a broad portfolio of underperforming models.
So Amazon leaned into what it does best: infrastructure. AWS is the landlord for much of the AI industry. The company has secured massive compute commitments, worth $138 billion from OpenAI and more than $100 billion from Anthropic. These deals underscore the growing demand for cloud infrastructure to train and run large language models. Amazon’s custom Trainium chips are now a multibillion-dollar business that Jeff Bezos has described as a fourth pillar of the company, alongside its e-commerce, cloud, and advertising businesses. The chips offer a more cost-effective alternative to NVIDIA GPUs for certain workloads, and they are increasingly attractive to AI companies looking to control their compute costs.
An Amazon spokesperson rejected the idea of a retreat. “AI models remain one of the most important things we’re working on, and that hasn’t changed,” the spokesperson said, adding that the company “continually evolve[s]” its lineup around what customers need. This statement reflects a desire to frame the changes as routine product evolution rather than a strategic failure. However, the internal realities suggest a significant downsizing of ambition.
The harder question is the one Nova never answered. Amazon can host everyone else’s models and sell the chips underneath them. That is a profitable and strategically important business. But whether Amazon can also build a frontier model that developers actively choose over Claude, Gemini, or GPT remains uncertain. The upcoming re:Invent conference will be a test of whether Amazon can deliver a model that is not just on par with its rivals, but genuinely preferred by the developer community.
The history of Amazon’s AI efforts is instructive. The company has long been a leader in applied AI, particularly in areas like recommendation systems, voice assistants, and supply chain optimization. But in the generative AI boom that began with the release of ChatGPT, Amazon appeared to be caught off guard. Its initial response was to offer access to third-party models through AWS Bedrock, rather than relying solely on its own models. This was a pragmatic approach, but it also signaled that Amazon was not confident in its ability to compete on model quality.
The decision to wind down most Nova models is a recognition of this reality. Rather than spreading resources across multiple models that struggle to gain traction, Amazon is concentrating its efforts on a single, potentially industry-leading model. This is a classic strategy of focus: do one thing well rather than many things mediocrely. The success of this strategy will depend on the quality of the new model and whether it can offer something unique that developers cannot get from existing frontier models.
Pieter Abbeel’s appointment to lead FMR is significant. Abbeel is a renowned researcher in robotics and reinforcement learning, with a track record of pioneering work in areas like imitation learning and robotic manipulation. His background suggests that Amazon may be aiming to build a model that is not just a chatbot, but a system capable of reasoning and acting in the physical world. This could differentiate Nova from the pure language models offered by competitors, though it is also a much harder technical challenge.
The competitive landscape has also shifted dramatically. OpenAI, Anthropic, Google, and Meta have all released increasingly powerful models, with capabilities in text, image, video, and audio generation. The cost of training these models is astronomical, and only a few organizations have the resources to compete. Amazon is certainly one of them, but its previous missteps have put it at a disadvantage. The company’s strength in cloud infrastructure and custom silicon gives it a unique vantage point, but it has yet to translate that into a leadership position in model development.
Another factor is the role of open-source models. The rise of capable open-weight models like Llama, Mistral, and DeepSeek has eroded the premium pricing power of proprietary models. This has forced even the largest companies to justify their model prices with superior performance or unique features. Amazon’s Nova models were seen as average in quality and not compelling enough to justify their cost. The new consolidated effort aims to change that perception.
The broader implications for AWS are also worth considering. Amazon’s cloud business benefits from being seen as a neutral infrastructure provider, offering access to all models, whether they come from OpenAI, Anthropic, or Amazon itself. However, if Amazon can build a truly competitive frontier model, it could create a more integrated stack that attracts customers to both AWS and its own AI services. This would mimic the strategy of Microsoft, which uses OpenAI’s models to enhance its Azure and Office products, while also investing heavily in its own AI research.
In the coming months, all eyes will be on re:Invent. If Amazon unveils a model that benchmarks competitively with Claude or GPT, it will validate the decision to consolidate. If the model underwhelms, it will raise serious questions about Amazon’s ability to ever catch up in the frontier race. Either way, Amazon’s strategy has shifted from breadth to depth, from having many models to having one great model. The company is clearly betting that focus will yield what sprawl could not: a place at the top of the AI stack.
Source:TNW | Amazon News