After nearly three years of intensive in-house development of cutting-edge large language models, Amazon has successfully launched its own model products. These models serve as cost-effective alternatives to flagship models from competitors, offering a balance of reliability and affordability, though they are not yet capable of complete substitution.
This reality is also reflected within Amazon's own product ecosystem: the Amazon shopping assistant Rufus operates on a hybrid architecture powered by both the self-developed Nova model and models developed by Anthropic; Amazon's emotionally intelligent programming tool Kiro, launched in mid-2025, currently incorporates models from Anthropic but has not yet integrated the Nova model. This reliance on external corporate models is raising concerns among some insiders at Amazon.
According to two individuals formerly involved in Amazon's AI work, leadership at Amazon Web Services has expressed concerns that Amazon's commercial applications built on Anthropic's technology risk becoming commoditized, making it difficult to establish a differentiated advantage. These sources indicated that management had initially hoped to develop more products using the Nova model to build unique competitive strengths for AWS offerings. However, engineers have so far been unable to overcome technical hurdles to bring the Nova model's performance to parity with Anthropic's models.
Due to these reasons, the sources said, AWS leadership has directed employees to invest more time in conducting post-training on the Nova models used in products, striving to further accentuate the differentiated characteristics of Amazon's offerings. Post-training is a technical process for optimizing and refining a model for specific tasks.
This situation highlights the dilemma facing Amazon's management: they must achieve technological independence in AI while adhering to the company's long-standing principle of financial prudence. Unlike competitors such as Meta Platforms, Amazon has consistently been unwilling to spend heavily to attract potential talent, which has somewhat limited its talent acquisition scale.
Amazon's AI team originated from the original Alexa smart assistant team. According to a former member of that team, Amazon's compensation structure, based on a "relatively low cash salary + long-term equity incentives" model, combined with a mandatory return-to-office policy, has placed it at a disadvantage in the fierce competition for AI talent. Two other individuals familiar with the AI business added that even within Amazon, some employees jokingly refer to these in-house models as "Amazon Basics," a playful nod to the company's private-label retail brand that carries a hint of irony.
An Amazon spokesperson responded, stating that claims about their compensation being "low-level" or "not top-tier competitive" neither accurately reflect their actual pay structure nor align with their investment and commitment to attracting and retaining world-class AI talent.
Indeed, growing evidence suggests that a "good enough" strategy for AI models can yield returns, with some companies opting for open-source models over the most advanced flagship models. However, if the gap with leading model providers becomes too wide, Amazon risks losing cloud business deals from customers deeply embedded in competitors' technology ecosystems.
The Amazon spokesperson stated, "Describing the relationship between the Nova model and other model vendors as a zero-sum game completely misrepresents our actual strategy. Our core objective is to provide AWS customers and Amazon's own businesses with the broadest possible range of AI choices. This means we are heavily investing in Nova model development while also deepening collaboration with partners."
The spokesperson added, "Far from distancing ourselves from partners, we are doubling down on cooperation. We just partnered with Anthropic to build the world's largest AI supercomputer, Project Rainier, powered by hundreds of thousands of our own Trainium2 training chips. The original intent behind developing the Nova model was to expand choice, not to restrict customers' technological paths."
Furthermore, Amazon stated that the differentiation of its AI products lies not in the underlying models, but in the customer experience built at the application layer and the deep integration capabilities with company data and infrastructure.
Judging solely by the key performance metric of AWS revenue growth, its performance has been a persistent concern for investors. Growth has hovered between 17% and 20%, significantly lower than the 30% to 40% growth rates of smaller rivals Microsoft Intelligent Cloud and Google Cloud. However, Amazon emphasized that its 20% year-over-year growth in Q3 2025 was the highest since 2022; considering its massive scale, such growth is quite impressive.
Overall, investors appear cautious about Amazon's AI strategy. In 2025, Amazon's stock price rose just over 5%, ranking last among major tech companies. Amazon's ability to meet the deployment needs of AI developers and related enterprises will significantly impact its cloud business development. Although Amazon established a partnership with Anthropic early on through investment and helping it expand its cloud customer base, Anthropic also maintains partnerships with Google and Microsoft.
Amazon stated that the latest version of the Nova model, publicly launched in December 2025, represents a major technological breakthrough compared to the initial version. According to an assessment by Artificial Analysis, the new Nova model outperforms comparable models from Anthropic and OpenAI in instruction following and agent tool use.
The Amazon spokesperson pointed out: "Tens of thousands of customers—including Cisco, Coinbase, Blue Origin, Reddit, Dentsu, and Infosys—are applying the Nova model to highly complex business scenarios. An Artificial Analysis evaluation report on the Nova2 model stated that 'Amazon is back among the world's top AI companies' and described the new model's performance as a 'quantum leap.' In 16 benchmark tests, the Nova2 Pro version matched the performance of Anthropic's flagship model in 10 tests, and met or exceeded GPT-5.1 levels in 8 others. Claims that customers must seek other models for reasoning needs completely ignore the fact that in most direct comparisons, the Nova2 model is mathematically comparable to, or even outperforms, competing models."
However, Artificial Analysis benchmarks show that the Nova2 model still ranks relatively low in areas like knowledge-oriented reasoning and agent programming. According to a person involved in Amazon's AI projects, this poses significant challenges for AWS employees developing the Kiro programming assistant and the enterprise search product Quick Suite using the Nova model. Currently, the Quick Suite utilizes a hybrid architecture powered by both Nova and Anthropic models.
In late December 2025, just weeks after Amazon released the Nova2 model, the company underwent a major reorganization of its AI teams. The AI division responsible for Nova model development was merged with quantum computing and chip development teams into a new organization led by AWS veteran Peter DeSantis. Rohit Prasad, who previously led the Nova project, is departing. Pieter Abbeel, founder of the AI robotics startup Covariant which joined Amazon in 2024, will take over advanced model research.
From Alexa to AGI Following the global frenzy triggered by OpenAI's ChatGPT launch in late 2022, Amazon established an independent AI lab in 2023 focused on LLM research. In the summer of 2023, under the leadership of Alexa Chief Scientist Prasad, Amazon transferred 8,000 employees from the Alexa team to a dedicated research unit named AGI, while also expanding the team through external hiring.
According to a former team member and another source familiar with Alexa, Alexa team members had extensive experience in natural language processing and machine learning, but relatively weaker expertise specifically in large language models.
One of the sources indicated that Amazon faced difficulties in external recruitment. On one hand, the direction of Amazon's AI team was significantly less attractive to talent compared to developing consumer-facing products like ChatGPT. On the other hand, two individuals who worked with the team stated there was a widespread perception that Amazon was not at the forefront of AI research.
Reportedly, in 2024, Amazon spent over $300 million to license technology from the AI agent-focused startup Adept and recruit its management talent. However, according to LinkedIn profiles, more than a dozen of the Adept employees who joined Amazon have since left for companies like Meta Platforms, Anthropic, and OpenAI.
The Amazon spokesperson stated: "For idealistic practitioners looking to create value for customers, Amazon is the best choice. Our researchers have direct access to world-class computing resources, including GPUs and Trainium chips, leveraging AWS's global infrastructure to apply technology to real-world business scenarios spanning Alexa, retail, logistics, and enterprise services."
The Deployment Journey of the Nova Model In December 2024, Amazon first publicly released the initial Nova model. Although its intelligence lagged behind competing models in industry benchmarks, its lower cost for enterprise customers won favor, primarily for handling basic business tasks.
For example, Nirmal Mukhi, Chief Architect at customer service technology company Asapp, said the company uses OpenAI's models to formulate solutions, employs Anthropic's models to retrieve existing data and generate optimal responses, while the first-generation Nova model is used for relatively simpler tasks like summarizing conversations.
He said, "Currently, we primarily use the Nova model for these simpler, lower-difficulty流程 tasks."
In early 2025, as AI agent technology gained market traction, Amazon launched the Nova Act service. This AWS-hosted service helps businesses build custom agents based on the Nova model. Customers generally view it as a reliable alternative, but not a cutting-edge technical solution.
AI startup Sola, which helps automate tasks like invoice processing, currently uses the Nova Act service based on the second-generation Nova model. Sola CEO Jessica Wu said Amazon promotes Nova Act and the underlying Nova model by emphasizing ultra-high reliability—the ability to run consistently and stably. Sola found Nova Act excels at automating repetitive manual processes and browser tasks.
She said, "This model is positioned as the most reliable enterprise-grade model. You can trust it not to crash easily, something almost no other foundational model can guarantee."
Throughout 2025, AWS launched several AI products for enterprise customers. During this time, Amazon employees internally attempted to integrate the Nova model into as many products as possible, but these efforts did not always yield the desired results.
For instance, according to a person involved in Amazon's AI projects, when developing the AI-powered enterprise search product Quick Suite, the team found the Nova model's performance in summarizing meeting minutes and refining information was inferior to other competitors, and the generated answers seemed rigid and mechanical.
Even newer versions released several months before Nova 2 had similar issues. Artificial Analysis evaluations showed the Nova 2 model's capability in knowledge-oriented reasoning remained weaker than comparable models—a capability precisely what engineers developing Quick Suite needed.
The Amazon spokesperson stated that Quick Suite, publicly launched in October 2025, indeed integrates multiple models, including Amazon's Nova.
The source also mentioned that applying the Nova model to programming assistant products like Kiro still faces technical challenges. Artificial Analysis benchmark results indicate that in agent programming capability, the second-generation Nova model still lags behind competing products from Anthropic and OpenAI.
Currently, Amazon's shopping assistant Rufus still operates on a hybrid architecture powered by both Nova and Anthropic models. Meanwhile, the Amazon spokesperson added, the Nova model currently provides core computing power for Amazon seller tools and the independent "Buy for Me" shopping feature launched in early 2025; Amazon's satellite internet service, Amazon Leo, also uses the Nova model to automate test case generation.
