Amazon, OpenAI, and Anthropic: AWS Competes on All AI Fronts

⚡

Key Takeaways

1Amazon has invested $50 billion in OpenAI and $8 billion in Anthropic, while developing its own AI models.

2AWS offers a "model-routing" service that directs requests to the most suitable model, strengthening its position as a central distributor.

3By developing its Trainium chips, Amazon reduces its dependence on NVIDIA, while integrating its Nova models into workflows via Bedrock.

💡Why it matters — This multifaceted strategy from Amazon could redefine competition and partnership dynamics in the AI sector.

Amazon Invests Heavily in AI

Amazon has recently made significant strategic decisions by investing colossal sums in two rival companies specializing in artificial intelligence, while also developing its own models. Matt Garman, the head of AWS, explained this approach at the HumanX conference in San Francisco. In February, Amazon injected $50 billion into OpenAI, after already investing $8 billion in Anthropic.

Garman emphasized that AWS's strategy has long been based on funding partners that it can then compete with. This model is illustrated by examples such as Oracle, which sells its databases through AWS, and Netflix, which uses AWS servers while building its own streaming infrastructure. This model is not new, but the scale of current investments is unprecedented.

AWS's "Model Routing"

AWS has implemented an innovative "model routing" service. This system aims to automatically direct customer requests to the most appropriate AI model for each specific task. For example, Claude is used for reasoning, GPT for planning, and a more economical model for code completion. According to Garman, this multi-model approach is set to become the norm, with AWS positioning itself as the central distributor rather than as the provider of a single model.

A Toll Strategy Inspired by the App Store

AWS's strategy resembles that of Apple's App Store, which takes a commission on each app without creating all of them. AWS aims to become the go-to toll for artificial intelligence. Before the February agreement, OpenAI and Anthropic's models were available on Microsoft Azure, AWS's main competitor. Failing to secure OpenAI would have meant ceding ground to Microsoft in the enterprise sector.

Internal Development and Self-Sufficiency

A less visible aspect of Amazon's strategy is the development of its own Trainium chips, aimed at reducing its dependence on NVIDIA. OpenAI has committed to deploying 2 gigawatts of Trainium capacity as part of their agreement, and Anthropic is already using these chips to train its models. Thus, every dollar invested in a competitor partially returns in the form of orders for Amazon's in-house silicon.

At the same time, Amazon is developing its own AI models, called Nova, which will be launched in late 2024 via the Bedrock platform. Although less publicized than Claude or GPT, these models hold an economic place in AWS's catalog. Thanks to model routing, AWS can discreetly integrate Nova into customer workflows, between two requests directed at third-party models. Ultimately, every company using Bedrock could consume Nova without being fully aware of it.

An Evolving Dynamic

This phenomenon of overlap between partners, competitors, and customers is not just a theory. During Anthropic's fundraising in February, at least a dozen investors from OpenAI also participated. Microsoft itself invested before announcing the integration of Claude into its products. The boundaries between partners, competitors, and customers are becoming blurred as investments increase.

For European companies, this situation goes beyond mere strategic curiosity. For example, Bouygues is already using Claude via Amazon Bedrock for contract analysis at its subsidiary Equans, and TF1+ operates on AWS. When the distributor also has its own models and chips, the promise of neutrality comes with a trust cost that each customer will need to evaluate.