Amazon is winding down most of its flagship Nova AI models as part of a strategic pivot to concentrate resources on a single, more ambitious frontier model. The company has begun deprecating several in-house Nova products, including the high-end Premier and Omni models, the Reel video generator, and the Canvas image generator. According to internal reports, some staff refer to this phase as "KTLO" — keep the lights on — meaning the models will remain supported for existing customers but will no longer receive significant development investment.
This is the concrete update to a story that first surfaced months ago, when Amazon shut its AGI Lab and signaled a retreat from the frontier AI race it never led. At the time, the news was centered on a closed research lab and accompanying layoffs. Now it is clear which products are being sacrificed, and what they are being sacrificed for.
One big bet instead of many
Resources are shifting to a new effort 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 research community. He is also a professor at the University of California, Berkeley, known for his work in reinforcement learning and robot manipulation. His leadership of FMR signals Amazon's intent to bring a more research-driven approach to its remaining AI ambitions.
A new flagship model is expected to debut at Amazon's re:Invent conference this autumn, according to reports. It could still carry the Nova name, which would preserve some continuity with the existing product line. The decision to consolidate around one model marks a significant departure from the previous strategy of spreading resources across multiple model types and modalities.
A shift in leadership and strategy
The logic behind the move is focus. Under previous AI chief Rohit Prasad, Amazon spread itself across text, image, and video models, attempting to cover every corner of the generative AI market. Prasad had championed the Nova family as a broad portfolio that could serve enterprise customers with different needs. But that approach proved difficult to sustain, especially as costs mounted and competition intensified.
Peter DeSantis, who took over the consolidated AI group in December, has concentrated talent and scarce compute resources on fewer frontier bets. DeSantis, a long-time Amazon executive known for his work on AWS infrastructure, has a reputation for operational discipline and a willingness to make hard trade-offs. His leadership has already produced visible changes: Prasad left at the end of 2025, and AGI Lab founder David Luan departed in February. These departures, combined with the closure of the San Francisco AGI site, an 80-person research group, underscore the scale of the reorganization.
The shift is not just about organizational structure. It reflects a broader realization that building a frontier AI model requires an enormous concentration of computing power, data, and research talent. Amazon's previous approach, which produced multiple models in parallel, was not delivering results that could compete with the likes of OpenAI, Anthropic, or Google. By consolidating, Amazon hopes to improve its odds of producing a single model that genuinely stands out.
What remains of Nova
Nova is not vanishing entirely. Amazon says several products will stay. The company is keeping Nova 2 Lite, Nova 2 Sonic, the Nova Forge customization service, and the Nova Act agent tool. These products serve specific enterprise use cases, such as lightweight text generation, low-latency speech, custom model fine-tuning, and autonomous web-based tasks. They are likely to continue generating revenue and providing value to existing AWS customers, even as the broader Nova line is pared back.
The deprecation of Premier, Omni, Reel, and Canvas, however, represents a major retreat from the multimodal ambitions Amazon once held. Omni was designed to handle text, vision, and audio inputs; Reel was intended to compete in the AI video generation space; Canvas targeted image generation. All of these are now being pushed to maintenance mode, a clear sign that Amazon no longer sees them as strategic priorities.
Where Amazon actually wins
Amazon never made Nova a household name the way OpenAI, Anthropic, and Google did with their models. Its systems also proved expensive to run for the value they returned. Cheaper rivals piled on the pressure, part of a wider shift to cut-price models that has redefined the economics of generative AI. For many developers, the cost-performance ratio matters more than brand recognition, and Amazon's models struggled to compete on that metric.
So Amazon leaned into what it does best: infrastructure. AWS is the landlord for much of the industry, with compute commitments worth $138 billion from OpenAI and more than $100 billion from Anthropic. These strategic deals give Amazon a central role in the AI boom, even when its own models are not the ones being deployed. Amazon's custom Trainium chips are now a multibillion-dollar business that Jeff Bezos, the company's founder and executive chairman, has called a fourth company pillar, alongside retail, AWS, and advertising.
This infrastructure advantage provides a stable revenue base and a strategic hedge. Even if Amazon never builds a frontier model that captures widespread developer enthusiasm, it can still profit from hosting the models built by its rivals. The compute commitments from OpenAI and Anthropic are not just financial wins; they also give Amazon insight into the demands of large-scale AI workloads and the future direction of model development.
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 official stance is consistent with Amazon's broader messaging, which has emphasized that the changes are about prioritization rather than withdrawal.
The open question
The harder question is the one Nova never answered. Amazon can host everyone else's models and sell the chips underneath them. Whether it can also build a frontier model that developers actively choose over Claude, Gemini, or GPT is what re:Invent will test. The conference, scheduled for autumn, will be the first major public showcase of Amazon's consolidated AI strategy. If the new flagship model impresses, it could reshape perceptions of Amazon as an AI innovator. If it falls short, Amazon may increasingly be seen as a pure infrastructure provider, content to power the success of others.
The road ahead is not without precedent. Amazon has a history of entering technology categories late and winning through persistence. AWS itself started as a modest internal infrastructure project before becoming the dominant cloud platform. But frontier AI is a different kind of challenge. The pace of progress is relentless, and the competition includes some of the best-funded and most aggressive companies in the world. Amazon's decision to bet everything on one model is a gamble, but it may be the only way to stay relevant in a game where scale and focus are decisive.
For now, the details of the new model remain under wraps. Reports suggest it will leverage FMR's research strengths and Amazon's vast cloud infrastructure. Whether it will be positioned as a general-purpose model or aimed at specific enterprise workloads is still unclear. What is clear is that Amazon has made its choice: fewer bets, bigger bets, and a singular focus on catching up in the race it has yet to lead.
Source: TNW | Amazon News