Amazon AGI

Amazon AGI is the artificial general intelligence division of Amazon.com, established in 2023 and headquartered in Seattle, Washington. It develops the Nova family of foundation models, oversees Amazon's Trainium custom AI silicon, and integrates with the broader Amazon Web Services AI portfolio including Amazon Bedrock, the managed-model distribution platform. As of April 2026, the unified AGI organization is led by Peter DeSantis following the December 2025 departure of inaugural division head Rohit Prasad.

At a glance

• Founded: 2023 as a dedicated AGI organization within Amazon. Earlier AI work spans Amazon Alexa (launched 2014) and Amazon Web Services AI services going back more than a decade.
• Status: Subsidiary of Amazon.com, Inc. Not independently capitalized.
• Funding: Operates within Amazon's R&D and AWS budgets. Parent company Amazon.com has a market capitalization of approximately $2.0 trillion as of April 2026 and is investing more than $100 billion annually in capital expenditure across AWS infrastructure, AI training compute, and custom silicon.
• Lead: Peter DeSantis (since December 2025). DeSantis runs a unified group covering AGI, custom silicon (including Trainium), and quantum computing, reporting directly to Amazon CEO Andy Jassy.
• Other notable leadership: Pieter Abbeel (frontier model research team lead, joined 2025), Andy Jassy (Amazon CEO; sets overall AI direction), Rohit Prasad (former SVP of Amazon AGI, led 2023 through 2025).
• Open weights: None. The Nova and Titan families are closed weights.
• Flagship products and models: Nova Pro, Nova Premier, Nova Micro, Nova Lite, Nova Canvas (image), Nova Reel (video), Nova Sonic (audio), Nova Forge (customizable), Trainium silicon, Bedrock.

Origins

Amazon's AI history precedes the formal AGI division. Amazon Web Services launched its first managed AI services in the mid-2010s (Polly, Lex, Rekognition, Comprehend), Alexa launched as a consumer voice assistant alongside the Amazon Echo in 2014, and Amazon SageMaker launched as a developer-facing machine-learning platform in 2017. The Titan family of foundation models, launched in September 2023 as a closed-weights line on Bedrock, was Amazon's first large-language-model release at scale.

In 2023, Amazon established a dedicated AGI organization led by Rohit Prasad, who had been head scientist for Alexa since joining Amazon in 2013. The organization centralized model development across what had previously been separate Alexa and AWS AI teams. The Nova family of foundation models, unveiled in December 2024, was the first major output of the consolidated organization, and superseded the Titan line as Amazon's primary frontier-model brand.

The strategic relationship with Anthropic began with an initial $4 billion investment in 2023 and expanded to approximately $8 billion in commitments by 2024, plus a deeper compute partnership in which AWS is Anthropic's primary cloud provider and Trainium-based clusters provide a substantial share of Anthropic's training compute. Project Rainier, announced in 2024, is a massive Trainium-based supercomputer dedicated to Anthropic training workloads.

In December 2025, Rohit Prasad announced his departure from Amazon at the end of the year. Andy Jassy named Peter DeSantis, a 27-year Amazon veteran who had previously led AWS computing products including the custom-silicon program, to lead a reorganized and unified AGI, custom-silicon, and quantum-computing division. Pieter Abbeel, the Berkeley BAIR professor and co-founder of robotics foundation-model company Covariant, joined Amazon AGI in 2025 to lead the frontier-model research team.

Mission and strategy

Amazon AGI's stated mission, articulated by Andy Jassy in 2024, is to deliver "frontier AI capability at scale, with the cost-performance and reliability that AWS customers expect." The framing is more enterprise-pragmatic than the AGI mission statements at OpenAI, Anthropic, and Google DeepMind, and reflects Amazon's commercial orientation toward AWS customers.

The strategy combines four threads. First, in-house frontier-model development through the Nova family, providing AWS customers with Amazon-branded model options at competitive cost-performance. Second, the strategic partnership with Anthropic, which provides AWS customers with frontier-tier Claude capability while Amazon supplies Anthropic with compute and chip infrastructure. Third, custom AI silicon (Trainium for training, Inferentia for inference), which differentiates AWS economically from clouds running NVIDIA-only stacks. Fourth, Amazon Bedrock, the managed-model distribution platform that hosts Nova alongside Anthropic Claude, Meta Llama, Mistral AI, Cohere, AI21, and other third-party model providers.

The competitive premise is that AWS customers value optionality and cost-performance more than they value being on the absolute frontier of capability, and that Bedrock's multi-vendor model marketplace is a more durable position than a single-vendor frontier strategy. The Nova family is positioned as the price-performance default for AWS customers, with Anthropic Claude as the frontier-tier upgrade and other third-party models for specialized cases.

The 2025 reorganization under DeSantis and the elevation of frontier-model research under Pieter Abbeel reflect an explicit corporate goal of closing the capability gap to OpenAI, Anthropic, and Google DeepMind for in-house Nova models, while retaining the multi-vendor Bedrock strategy and the deepening Anthropic relationship.

Models and products

• Nova family. Amazon's primary foundation-model line, unveiled December 2024. Includes Nova Micro (text-only), Nova Lite (multimodal small), Nova Pro (multimodal mid-tier), and Nova Premier (multimodal flagship). Nova Canvas generates images; Nova Reel generates short-form video. Nova Sonic, released in March 2025, adds real-time audio capability. Nova Forge, released in December 2025, lets sophisticated customers fine-tune Nova variants on their own data. All closed weights.
• Titan family. Earlier closed-weights line including Titan Text Express (September 2023), Titan Image Generator (November 2023), and Titan Embeddings. Distribution continues on Bedrock alongside Nova.
• Amazon Bedrock. Managed-service platform for running Amazon's own models alongside third-party models including Anthropic Claude, Meta Llama, Mistral, Cohere, AI21, and Stability AI offerings. Bedrock is the primary AWS distribution channel for AI inference workloads.
• Bedrock Agent Core. Agent-development platform launched in 2025 for building tool-using agents with Bedrock-hosted models.
• Trainium silicon. Custom AI training accelerators. Trainium 2 launched in 2024 with substantial deployment in Project Rainier, the massive Trainium cluster dedicated to Anthropic training. Trainium 3 has been announced for future deployment.
• Inferentia silicon. Custom AI inference accelerators. Used to lower the per-token cost of inference workloads on AWS.
• Alexa+. AI-powered consumer assistant, the next generation of Amazon Alexa, integrating large-language-model capability into the existing Echo device base. Launched in 2024.
• Amazon SageMaker. Developer-facing machine-learning platform for custom model training and deployment.

Distribution channels include Amazon Bedrock for managed-service inference, AWS SageMaker for custom-model workflows, the Alexa product line for consumer-AI distribution, and Amazon's first-party retail and logistics applications for internal AI deployment.

Benchmarks and standing

Amazon does not consistently publish frontier benchmark comparisons for the Nova family, and the Nova line has not been comprehensively represented on Artificial Analysis Intelligence Index, LMArena, GPQA Diamond, or SWE-bench Verified leaderboards as of April 2026. Public Amazon materials emphasize cost-performance and integration into AWS workflows over standardized capability benchmarks.

Where third-party evaluation has been published, Nova Pro and Nova Premier have shown competitive performance on cost-performance metrics for enterprise workloads but have not matched the absolute capability of leading frontier models from OpenAI, Anthropic, and Google DeepMind. The strategic positioning of Nova as a price-performance default rather than a capability frontier reflects this gap.

The Nova benchmark position should be interpreted alongside Bedrock's multi-vendor strategy. AWS customers running Bedrock for frontier-tier capability typically select Anthropic Claude or other partner models rather than Nova; Nova is positioned for the cost-sensitive bulk of AWS AI inference workloads.

Leadership

As of April 2026, Amazon AGI's senior leadership includes:

• Peter DeSantis, leader of the unified AGI, custom-silicon, and quantum-computing division (since December 2025). 27-year Amazon veteran. Previously led AWS compute, storage, database, security, and custom-chip product teams. Reports directly to Andy Jassy.
• Pieter Abbeel, frontier-model research team lead (joined 2025). Berkeley BAIR professor and co-founder of Covariant, the robotics foundation-model company. Brings academic-research credibility to the in-house frontier-model program.
• Andy Jassy, CEO of Amazon.com Inc. Sets overall AI direction. Amazon's broad capital expenditure in AI infrastructure (Trainium, Project Rainier, the Anthropic partnership) reflects Jassy's strategic priorities.
• Adam Selipsky's successor, currently Matt Garman as CEO of Amazon Web Services since 2024. AWS revenue, including Bedrock and SageMaker, is the principal AI commercial channel.

The departures cohort includes Rohit Prasad, who led Amazon AGI from 2023 through 2025 and oversaw the creation of the Nova family. The leadership reorganization in December 2025 was unusual in scope, combining AGI, custom silicon, and quantum computing under a single senior leader, and reflects an explicit attempt to consolidate Amazon's AI infrastructure and model-development capabilities under unified strategy.

Funding and backers

Amazon AGI does not raise independent capital. The organization operates within Amazon.com Inc.'s R&D and AWS budgets. Amazon's parent company has a market capitalization of approximately $2.0 trillion as of April 2026 and is the second-largest cloud provider globally.

The most significant external AI-related capital flow is the Anthropic partnership. Amazon's cumulative commitment to Anthropic reached approximately $8 billion through 2024, structured as a combination of equity investment and AWS compute commitments. The arrangement makes Anthropic Amazon's largest external AI investment, more than offsetting Amazon's relatively modest acquisition activity in the AI startup space.

Amazon's broader AI capital expenditure has expanded substantially. Public AWS guidance through 2025 and 2026 has indicated tens of billions of dollars in annual capital expenditure on AI data centers, Trainium and Inferentia silicon production, and the Project Rainier compute infrastructure dedicated to Anthropic. Total Amazon capital expenditure is projected at more than $100 billion annually, of which AI infrastructure is the fastest-growing component.

Industry position

Amazon AGI occupies a structurally distinctive position among Incumbent AI labs. The combination of in-house frontier-model development (Nova), the deep partnership with Anthropic including Project Rainier compute, the custom Trainium and Inferentia silicon program, and Bedrock's multi-vendor distribution platform produces a strategic profile no other Incumbent has matched. Amazon's AI strategy explicitly leverages the company's existing AWS distribution moat rather than competing primarily on raw model capability.

Strategic risks identified in industry coverage include the in-house frontier-capability gap to OpenAI, Anthropic, and Google DeepMind (which constrains Nova's positioning to cost-performance rather than capability leadership), the structural tension between investing in in-house Nova and supplying compute to Anthropic Claude (which competes with Nova on AWS customer wallets), and execution risk of the December 2025 reorganization combining AGI with custom silicon and quantum computing under DeSantis's expanded mandate.

The strategic strengths are equally distinct. AWS distribution at scale provides revenue from third-party model hosting (Anthropic, Meta, Mistral, Cohere) regardless of whether Amazon's own Nova line wins on capability. The Trainium silicon program reduces unit economics of inference and training relative to NVIDIA-only competitors. The Anthropic partnership provides Amazon with a frontier-capability hedge that requires no in-house commercialization risk.

Competitive landscape

Amazon AGI competes with several Frontier and Incumbent labs:

• Microsoft AI. Direct competitor in cloud AI, where Azure AI Foundry competes with AWS Bedrock for enterprise model-hosting workloads. Microsoft's OpenAI partnership parallels Amazon's Anthropic partnership in shape.
• Google DeepMind. Cloud AI competitor through Google Cloud Vertex AI. Gemini in Vertex competes with Nova and Anthropic Claude in Bedrock.
• OpenAI. Frontier-capability competitor (Nova versus GPT-5.5) and increasingly an enterprise sales competitor as OpenAI pivots to enterprise in 2026.
• Anthropic. Strategic partner and de facto competitor on Bedrock, where Claude and Nova compete for AWS customer wallets even as Amazon supplies Anthropic with compute.
• Meta AI / FAIR. Less direct competition; Meta's open-weights Llama family is hosted on Bedrock as a third-party option.
• NVIDIA. Less a model-capability competitor and more a silicon competitor where Trainium and Inferentia compete with NVIDIA H100 and successors for cloud AI workloads.

Outlook

Several open questions affect Amazon AGI's trajectory in 2026 and 2027:

• The capability progression of the Nova family relative to OpenAI, Anthropic, and Google DeepMind frontier lines, and whether Nova can move from cost-performance positioning toward capability competition.
• The execution of Peter DeSantis's expanded mandate covering AGI, custom silicon, and quantum computing, including organizational integration and product roadmap coherence.
• The trajectory of the Anthropic partnership, including any expansion of Amazon's investment, deepening of the Project Rainier compute commitment, or commercial implications for Bedrock if the relationship terms shift.
• Trainium 3 deployment and adoption, and whether AWS can defend silicon margins against NVIDIA's continued capability advances.
• Continued commercial traction for Alexa+ in the consumer market and the role of in-house Nova versus third-party models in Alexa+ generations.
• Pieter Abbeel's research output and any frontier-tier capability claims for in-house Nova models that close the gap to leading frontier labs.

Sources

• CNBC: Amazon says AI chief Rohit Prasad is leaving, Peter DeSantis to lead 'AGI' group. December 2025 leadership announcement.
• GeekWire: Amazon AI chief Rohit Prasad leaving; Peter DeSantis to lead unified AI group. Reorganization context.
• About Amazon: Andy Jassy makes Amazon leadership announcement. Official leadership announcement.
• About Amazon: Introducing Amazon Nova. December 2024 Nova launch.
• Constellation Research: AWS launches Bedrock Agent Core, custom Nova models. Bedrock and Nova Forge context.
• AWS: Generative foundation model, Amazon Nova. Official Nova product page.
• Wikipedia: Amazon Web Services. Reference material.