Amazon Shrinks Nova Model, Focuses Resources on Frontier Model Program Led by Pieter Abbeel

Historically, Amazon's layout in the large model and agent direction has undergone multiple rounds of organizational and strategic shifts: initially focused on Alexa and cloud models, it later established the AGI Lab through "acquisition-based hiring" of the Adept team in 2024, attempting to accelerate its catch-up in general agents and AGI; however, within two years, the AGI Lab was closed, its leaders left, and recently there have been layoffs within the AGI organization, consolidating related work under a larger technology department. This indicates a shift from "exploring AGI through multiple lines" to a contraction adjustment of "reducing trial and error, focusing on a few frontier models."

From a capital perspective, the Nova series represents Amazon's "productized model portfolio" aimed at large clients and developers: Premier and Omni compete with general large models, while Reel and Canvas provide video and image generation capabilities. Transferring the high-end text, video, and image generation models to a "maintain operations, no longer focus on development" status essentially treats the previously locked-in revenue and customer relationships as cash flow assets, while all new engineering and computing budgets are redirected to frontier model research—this is essentially a path of "cutting tail product lines and concentrating capital expenditure on a single flagship bet," freeing up computing power and budget space for potentially stronger foundational models in the future.

In horizontal comparison, OpenAI has also gradually narrowed its product line to GPT-4/4.1 and specific capabilities after GPT-3/3.5, while Anthropic has concentrated resources around the Claude 3 series, no longer maintaining too many old versions; what Amazon is doing now is a latecomer's version: first building a Nova product family, then realizing that it still lags behind in top-level foundational models and reasoning capabilities, thus pausing the development of some flagship application models and concentrating engineering and computing resources on the frontier model project led by Pieter Abbeel, hoping to reverse its competitive position in the cloud AI and foundational model market through a "super frontier" model—this reflects a shift from "broadly spreading product lines" to "concentrating bets on foundational models and frontier research," a typical strategy of a catch-up player.

Structurally, this adjustment belongs to "capital concentration + technological substitution": on the technology level, the new frontier model is expected to replace the roles of multiple current Nova text, video, and image models, supporting a broader product line with a single or few foundational models; on the capital level, computing budgets and top research talent are being pulled from multiple teams and concentrated on the EMR / FMR high-priority project, indicating that Amazon is willing to give up some short-term product diversity and mid-tier customer customization space in exchange for a flagship model that has the opportunity to compete with top competitors in parameter scale, reasoning capability, and multimodal performance—whoever masters such frontier models will control the pricing power and ecological dominance of cloud AI services. By shrinking the Nova array and closing the AGI Lab, Amazon is restructuring resources into a "high-risk, high-potential-return" concentrated bet, attempting to gain a structural position in the next round of foundational model reshuffling rather than maintaining a long-tail supplier role.

ABAB News · Cognitive Laws

1. The more product lines, the safer; the more expensive computing power, the more need for concentrated bets.

2. If a follower does not cut old models, it can only forever lose to the leader in computing power.

3. Tools can be diversified, but foundational models can only be represented by a single flagship.