Amazon AWS: AI Dependency on OpenAI and Anthropic

Key Takeaways

• Amazon invested approximately $50 billion in OpenAI (April 2026) and up to $33 billion in Anthropic, directing over $80 billion to external AI partners.
• In July 2026, Amazon discontinued most of its own Nova models, including Nova Premier, Omni, Reel, and Canvas – they are in "Keep the Lights On" status with no active development.
• Anthropic is contractually obligated to run AI workloads primarily on AWS and use Amazon's Trainium and Graviton chips, supporting the company's hardware strategy.
• A significant portion of AWS's AI revenue currently depends on OpenAI and Anthropic, creating contractual and strategic dependencies.
• Microsoft holds a structural advantage over AWS with its close OpenAI partnership, as OpenAI models are deeply embedded in Azure and the Microsoft ecosystem.
• AWS CEO Matt Garman defended the parallel investments in both AI labs as part of the strategic DNA of cloud business – working with partners while competing with them.

Billion-Dollar Bets on External AI Partners

Amazon Web Services (AWS) currently monetizes artificial intelligence primarily through infrastructure services and AI workloads. In spring 2026, the company pursued a dual partnership strategy: In April 2026, Amazon invested approximately $50 billion in OpenAI, after OpenAI committed to spending $100 billion on AWS infrastructure and softened its previous Microsoft exclusivity. By late April 2026, OpenAI models such as GPT-5.5, GPT-5.4, and GPT-5.6 were made available via the AWS platform Bedrock.

In parallel, Amazon deepened its engagement with Anthropic, the developer of Claude models. Following an initial investment of about $8 billion, an additional investment of up to $25 billion was announced in April 2026. The total package with Anthropic amounts to approximately $33 billion as of April 22, 2026. In return, Anthropic commits to running AI workloads primarily on AWS and using Amazon's own AI chips Trainium and Graviton – a contractual arrangement that helps Amazon reduce its dependence on external chip manufacturers.

Strategic Justification for Parallel Investments

AWS CEO Matt Garman defended the parallel engagements with OpenAI and Anthropic in April 2026 as an integral part of cloud strategy. The approach of working with partners while competing with them has been part of AWS's business model since its early days. Garman emphasized that AWS is careful not to give its own offerings unfair advantages over partners, even when there are overlaps in the product portfolio.

The OpenAI engagement is also a response to competition with Microsoft. Both AI models – from OpenAI and Anthropic – were previously available via Microsoft's Azure cloud, putting AWS under pressure. The OpenAI deal reduces dependency on the previous primary partner Anthropic and makes AWS more attractive to large enterprise customers seeking access to multiple leading foundation models. A central market trend is so-called model routing – systems automatically select the most suitable AI model for a given task.

Withdrawal from Own Model Development

In July 2026, Amazon underwent a strategic reversal in its own AI development. The company discontinued most of its self-developed Nova models and concentrated developers and computational resources on a single frontier project. Specifically, the high-end models Nova Premier, the multimodal Omni model, the video model Reel, and the image model Canvas are internally in "Keep the Lights On" status – they continue to be operated for existing customers but are no longer actively developed.

The remaining portfolio includes Nova 2 Sonic, Nova 2 Lite, Nova Forge (a service for customers to build their own models), and agent technology Nova Act. The restructuring followed a round of layoffs in the Artificial General Intelligence (AGI) organization and the closure of the AGI Lab. The lab was established in 2024 following Amazon's acquisition of AI startup Adept and worked on long-term research. Adept co-founder David Luan left Amazon in February 2026, after which the lab was closed and the San Francisco location was abandoned.

In parallel within the AGI organization, the Frontier Model Research (FMR) group was established, led by UC Berkeley professor Pieter Abbeel, who joined Amazon through the acquisition of robotics startup Covariant. FMR is considered the top priority internally, with resources shifting away from existing Nova models. According to internal reports from July 2026, the new foundation model being developed there is set to debut at Amazon's re:Invent conference in the fall.

Structural Dependencies and Competitive Risks

With the withdrawal from broad own model development, a structural dependency emerges: A significant portion of AWS's AI revenue currently depends on OpenAI and Anthropic. A material share of expected AI earnings is tied to the continuation, intensity, and terms of these partnerships. Changes in contractual arrangements, strategic shifts by partners, or moving away from AWS could noticeably impact the revenue base.

The competitive landscape versus Microsoft remains asymmetrical. Microsoft holds a structural advantage with its close, exclusively structured partnership with OpenAI. OpenAI models are deeply embedded in Azure and Microsoft's product ecosystem – from Office 365 to GitHub Copilot to Windows features. This gives Microsoft not only differentiation in the cloud market but also privileged access to OpenAI's technical innovations. For Amazon, this creates a disadvantage: While AWS is equally strong in the AI infrastructure business, it lacks comparable exclusive rights to a dominant foundation model provider. This makes it harder to clearly differentiate itself from Azure in the high-margin segment of specialized AI services.

Focus on Infrastructure Rather Than Proprietary Models

The restructuring represents a shift in Amazon's AI strategy: As own foundation models recede into the background, the company is increasingly focusing on providing infrastructure – GPU clusters, computing capacity, security and control mechanisms such as IAM, PrivateLink, Guardrails, and CloudTrail. Enterprise customers should be able to access models from OpenAI and Anthropic via AWS, embedded in Amazon's security architecture.

An Amazon spokesperson emphasized in late July 2026 that AI models are traditionally supported over long periods because customers depend on them, and that the company remains committed to investments in frontier models. The model portfolio is continuously adjusted to customer needs. The contractually secured use of Amazon's Trainium and Graviton chips by Anthropic further supports the hardware strategy and reduces dependence on Nvidia.

For investors, the picture is mixed: AWS benefits from rising demand for AI infrastructure and secures revenue streams through contractual commitments. At the same time, the company lacks a proprietary flagship model that could ensure long-term differentiation and pricing power. The dependence on external partners – particularly on OpenAI, which remains closely tied to Microsoft – remains a structural risk in a rapidly changing market.