Amazon has quietly pulled the plug on active development for most of its flagship Nova AI models, according to a Business Insider report from July 28. The company is winding down Nova Premier, Nova Omni, its Reel video generation model, and Canvas image generation tool in a sweeping overhaul of its artificial intelligence strategy.

For businesses building on AWS Bedrock, this is a signal worth paying attention to.

What Amazon Is Actually Doing

The models are not being deleted overnight. Amazon is moving them into what it internally calls “KTLO” mode, short for “keep the lights on.” Existing customers can keep using them, but the models will no longer receive new features, performance improvements, or active investment.

The resources being freed up are going to a new internal initiative called Frontier Model Research, led by Pieter Abbeel, the AI researcher who joined Amazon through its 2024 acquisition of robotics startup Covariant. The goal is to build a single, competitive frontier-scale model rather than maintaining a sprawling portfolio of specialised tools.

Amazon is targeting its re:Invent conference later in 2026 as the debut window for this new flagship model. The Nova brand may survive on the new system.

Not everything in the Nova family is being shelved. Nova 2 Lite, Nova 2 Sonic (for voice applications), and Nova Forge, which lets customers build customised models using Amazon’s technology, are continuing. So this is less a complete withdrawal and more a sharp consolidation.

Why This Matters

Amazon launched the original Nova family at re:Invent 2024 with a lot of fanfare, positioning it as a multimodal workhorse for the Bedrock platform. Less than two years later, the strategy is shifting.

The move reflects a reality that has been building across the industry: running parallel model families is enormously expensive, and the gap between good and great in foundation models is widening. Instead of maintaining several models that compete in different capability brackets, Amazon is making the same bet OpenAI, Anthropic, and Google have been making all along. Build one extremely capable model and iterate on it.

This matters for the competitive picture too. Amazon’s existing partner Anthropic provides Claude models through Bedrock, and businesses have increasingly gravitated toward those for reasoning and agentic use cases. The Nova Premier and Omni models were supposed to give Amazon more control over its AI stack, but that plan is being reconsidered.

The leadership dimension is also notable. Amazon CEO Andy Jassy recently announced the departure of AI exec Rohit Prasad, who had overseen much of the Nova development. Abbeel represents a different kind of AI leadership, one more focused on research-forward frontier capabilities than applied product development.

What This Means for Business

If you are using Nova Premier or Omni on AWS Bedrock today, your workflows are not immediately broken. Amazon has committed to keeping these models accessible for existing users. But you should start planning for eventual migration, because “keep the lights on” has a shelf life.

If you were evaluating Nova models for new projects, this announcement is a clear signal to look elsewhere in the near term. Anthropic’s Claude models on Bedrock are the more obvious choice for reasoning and agentic tasks. For image and video generation, third-party providers remain active options.

The broader lesson is about vendor AI strategy. Companies betting on proprietary in-house models face enormous resource requirements to stay competitive. Amazon’s pivot shows that even the largest cloud providers are recalibrating. The frontier model race is increasingly a three-horse contest between OpenAI, Anthropic, and Google, with everyone else either partnering or buying in.

For organisations building AI systems on cloud infrastructure, this reinforces the importance of building on abstraction layers rather than locking into specific model APIs. The underlying models will keep changing. What matters is whether your workflows can adapt when they do.

Amazon’s bet is that one great model beats several average ones. By the time re:Invent arrives, we will know whether that strategy holds.