Rohit Prasad leaves Amazon AI
Rohit Prasad's departure from Amazon at the end of the year is part of a comprehensive reorganization. Amazon is bundling AI models, custom chips, and quantum projects under Peter DeSantis. This realignment signals a strategic shift towards an integrated 'stack' thinking that encompasses silicon to models.
Source: Proprietary Depiction
Amazon's AI Strategy
Amazon CEO Andy Jassy confirmed that Rohit Prasad will leave Amazon at the end of the year. This decision is viewed by Reuters and Business Insider as part of a larger restructuring within Amazon's AI and cloud divisions. Prasad joined Amazon in 2013 to build Alexa, , which today reaches hundreds of millions of customers daily. . For the past two years, he led the development of Amazon Nova and the internal 'AGI' organization, which produced twelve foundation models, , used by tens of thousands of businesses. This development positioned Prasad as the chief engineer for a proprietary model portfolio, as reported by Reuters. . The term 'AGI' remains Amazon's internal designation and is primarily visible through memos and reporting, as The Verge notes.
Source: americanbazaaronline.com
Rohit Prasad, the outgoing head of Amazon AI.
Jassy describes the current timing as an "inflection point" for new technologies. Peter DeSantis will lead a new organization that encompasses comprehensive AI models (including Nova and 'AGI'), silicon development (Graviton, Trainium, Nitro), and quantum computing encompasses comprehensive AI models (including Nova and 'AGI'), silicon development (Graviton, Trainium, Nitro), and quantum computing . DeSantis will report directly to Jassy, underscoring the company's priority in this area. Reuters and the Financial Times confirm this broader realignment. DeSantis has been with Amazon for over 27 years, was involved in the launch of EC2, and oversaw the infrastructure and utility computing world, which spans over 38 regions and 120 Availability Zones . He drove the acquisition of Annapurna Labs, in 2015, which gave rise to Amazon's custom silicon team.
Source: adgully.com
Rohit Prasad, a key figure in the development of Amazon Alexa.
Pieter Abbeel will lead the 'frontier model research team' and continue to collaborate with the robotics team, as Reuters. Amazon had hired Abbeel and other Covariant founders in 2024 and licensed Covariant's robotics foundation models, , which TechCrunch beschreibt. Abbeel is thus part of a longer-term 'Robotics + Foundation Models' strategy with a clear research mandate at the core of the model organization.
Technological Pillars
Jassy points to the benefits of co-optimization "across models, chips, and cloud software and infrastructure." The Nova-2 models were unveiled at re:Invent and promise "reasoning capabilities" at "industry-leading price performance." AWS announced Nova 2 Lite as a fast, cost-effective model for everyday workloads and Nova 2 Pro (preview) for complex multi-step tasks. The Nova family is available in the AWS product catalog. . Technical reality for teams often runs through Amazon Bedrock, , where Amazon's proprietary models and partner offerings like Anthropic converge. AWS described Nova 2 Lite in more detail on the AWS News Blog , emphasizing its use for "agentic AI applications." Amazon is bundling re:Invent announcements, , including Bedrock-AgentCore, Trainium3 UltraServers, and new Graviton processors. re:Invent 2025 took place from November 30 to December 4 in Las Vegas.
Source: YouTube
Jassy lists Graviton, Trainium, and Nitro as examples of "silicon development" . AWS positions Graviton as a processor family, intended to deliver "best price performance" for EC2 workloads. Graviton-based EC2 instances can offer up to 40% better price-performance compared to comparable x86 instances. Trainium is AWS's own accelerator family for training and inference. In December 2025, AWS announced 'EC2 Trn3 UltraServers' , which are expected to offer Trainium3 with up to three times the performance and more "output tokens per megawatt." Nitro instances combine dedicated hardware and a lightweight hypervisor architecture that shapes the EC2 infrastructure.
Quantum computing is the third pillar of the new organization. Amazon Braket is a managed service for quantum computers and simulators. In March 2025, Amazon introduced its own quantum chip "Ocelot", , which is designed to make error correction more efficient, with a claimed reduction of up to 90%. . The coupling of quantum computing with AI and chips serves to centrally bundle long-term, computationally intensive platform themes in order to align infrastructure, roadmaps, and investment logic.
Source: YouTube
Market Context and Outlook
Amazon's AI strategy pursues a two-pronged approach: proprietary models and a marketplace via Bedrock. A key partner is Anthropic. Amazon is investing an additional $4 billion in Anthropic an additional $4 billion in Anthropic, , bringing the total investment to $8 billion and making AWS the primary cloud and training partner. Claude models are consumable via Bedrock, , and AWS lists supported Claude models for the Bedrock messages API . The DeSantis reorganization more tightly links Amazon's proprietary model program (Nova/"AGI") to infrastructure economics, while Bedrock retains breadth through partners. This is a division: differentiation through stack and efficiency, choice through platform.
It remains to be seen how quickly Nova 2 grows from "announcement + preview" into stable availability. AWS „What’s New“ and the Bedrock model documentation are indicators for this. Second, whether Amazon measurably backs up the price-performance narrative with its own chips. Trainium3 performance and efficiency claims Trainium3 performance and efficiency claims as well as reference architectures are the litmus test. Third, how the research line under Abbeel becomes visible, whether in publications, benchmarks, or collaborations, as mentioned in the mentioned. Fourth, what role quantum computing plays in practice. memo mentions. Braket Braket and Ocelot must translate into workloads, programs, and industry partnerships that go beyond pilot projects.
Rohit Prasad's departure at year-end is confirmed by the Amazon-Memo and media reports. . The more significant news is the new leadership structure under Peter DeSantis, which directly brings Nova/"AGI", custom silicon (Graviton/Trainium/Nitro), and quantum computing under the CEO. This is to be understood as "AI becomes platform engineering": optimization of costs, performance, infrastructure, and roadmap, parallel to a partner ecosystem that is visible through Bedrock and its documentation is visible.