Amazon to Invest Up to $4 Billion in Anthropic

Amazon announced today an investment of up to $4 billion in Anthropic, the AI startup founded in 2021 by Dario Amodei and Daniela Amodei, both former OpenAI research executives. The deal makes Amazon Web Services (AWS) Anthropic's primary cloud computing provider and puts Amazon's own chips, Trainium and Inferentia, at the center of training and running Claude, the company's large language model.

A staged payout, not a single check

According to the joint announcement from both companies, Amazon will initially put in $1.25 billion, with the option to expand the investment up to $4 billion depending on future milestones. It's a minority stake: Amazon won't take control of Anthropic or a decisive board seat, something the startup has guarded carefully since its founding to preserve independence over its safety and research decisions.

Money isn't the only piece of the deal. Anthropic commits to using AWS as its primary cloud provider for training and inference workloads, and both companies will collaborate on developing Amazon's future AI-specific chips. Trainium, built for training models, and Inferentia, designed to run them efficiently, have long been Amazon's bet on reducing its dependence on Nvidia GPUs, today the industry's most expensive and scarce bottleneck.

Claude, a GPT-4 rival that needs compute muscle

Anthropic launched Claude 2 this past July, a model that competes directly with OpenAI's GPT-4 and Google's Bard on conversation, summarization and text generation tasks. Training and serving models of this scale requires computing infrastructure few startups can afford on their own, and that's precisely the logic that has pushed major cloud providers to buy into generative AI labs: they don't just gain exposure to a booming technology, they also lock in a customer for their cloud and, in Amazon's case, for its own chips.

Anthropic already had Google's backing, which invested roughly $300 million back in February of this year in exchange for a minority stake in the company. Amazon's entry doesn't replace that deal — it adds a second major infrastructure backer, an unusual arrangement that reflects both the capital appetite required to train frontier models and Anthropic's determination not to depend on a single cloud provider.

The shadow of the Microsoft-OpenAI deal

Amazon's move is best understood against what Microsoft did earlier this year. In January, Microsoft announced an investment pegged at around $10 billion in OpenAI, expanding a relationship dating back to 2019 that made Azure the near-exclusive infrastructure for training and serving OpenAI's models, including GPT-4. That deal reshaped the competitive cloud landscape: every major OpenAI announcement also reinforced Azure's position against AWS and Google Cloud.

AWS has been the dominant cloud provider by market share for years, but it had arrived late to the conversation around proprietary foundation models compared with Microsoft and Google. Locking in Anthropic as a strategic customer and a test bed for its AI chips is Amazon's most direct way of staying in the race to control the infrastructure layer on which the next generation of AI applications will be built.

What changes from here

For AWS's enterprise customers, the deal points to deeper integration of Claude within Amazon's services, along lines similar to how Microsoft has woven OpenAI's models into Azure and its product suite. For Anthropic, it means the financial backing needed to compete in a race where training a single frontier model can cost tens of millions of dollars, without having to depend on a single partner for all its infrastructure.

It remains to be seen how the relationship between Amazon and Google evolves as simultaneous backers of the same startup, and whether Trainium and Inferentia can prove, through Anthropic's actual use of them, that they're a viable alternative to Nvidia GPUs. That real-world validation, more than the size of the check, will determine whether this investment ends up being decisive for AWS's future in AI.