Amazon.com Inc. will invest up to $4 billion in Anthropic, the San Francisco-based startup behind the Claude 2 large language model.

The companies announced the investment this morning. The deal comes as Anthropic reportedly gears up to develop a highly advanced foundation model, Claude Next, that could be too hardware-intensive for even the world’s fastest supercomputers. The capital infusion from AWS should make it easier for the company to shoulder the model’s likely steep development costs.

“We have tremendous respect for Anthropic’s team and foundation models, and believe we can help improve many customer experiences, short- and long-term, through our deeper collaboration,” said Amazon Chief Executive Officer Andy Jassy. 

Anthropic was launched in 2020 by former OpenAI LP researchers to build generative artificial intelligence software. Its newest and most advanced model, Claude 2, is designed to compete with OpenAI’s GPT-4. It can generate marketing copy, solve mathematical problems and turn natural language prompts into software code.

One of Claude 2’s flagship features is its ability to process prompts containing up to 100,000 tokens. A token is a unit of data in AI development that corresponds to a few characters or numbers. Because it can analyze large amounts of data, Claude 2 lends itself to tasks such as summarizing long legal documents that are difficult to perform with other language models.

As part of the investment announced today, Amazon will buy a minority stake of unspecified size in Anthropic for $1.25 billion. The terms of the deal give the cloud and online retail giant the option to eventually invest an additional $2.75 billion. The companies didn’t disclose the valuation at which the investment was made.

The funding round builds on a partnership that began in 2021, when Anthropic became a customer of Amazon’s Amazon Web Services Inc. cloud unit. Earlier this year, AWS made Claude 2 available through its Amazon Bedrock generative AI service. The companies plan to extend their partnership into yet more areas following today’s investment.

Going forward, Anthropic will run the majority of its workloads on AWS. The Amazon unit is set to become its “primary cloud provider for mission critical workloads,” including model development and AI safety projects. Anthropic will use instances powered by the cloud giant’s internally developed AWS Trainium and AWS Inferentia chips to support its research.

Trainium is a series of specialized processors, now in its second generation, that is optimized specifically to train AI models with more than 100 billion parameters. AWS enables companies to provision up to 16 such processors per instance. According to the cloud giant, Trainium processors can train AI models for a significantly lower cost than competing processors.

Anthropic also plans to use AWS’ Inferentia chips, which are optimized for inference. That’s the task of running already trained models in production. AWS says Inferentia chips can provide up to 230% higher throughput per inference than competing silicon while reducing costs by 70%.

Anthropic will not only use the cloud giant’s internally-developed AI chips but also help improve them. The companies said that they intend to “collaborate in the development of future Trainium and Inferentia technology.” 

AWS and Anthropic will expand their collaboration in the go-to-market arena, too. Anthropic has made a “long-term commitment” to make its future AI models available through AWS’ Bedrock service. Joint customers will receive early access to new features, including capabilities that can be used to customize AI models.

“Since announcing our support of Amazon Bedrock in April, Claude has seen significant organic adoption from AWS customers,” said Anthropic co-founder and CEO Dario Amodei. “By significantly expanding our partnership, we can unlock new possibilities for organizations of all sizes, as they deploy Anthropic’s safe, state-of-the-art AI systems together with AWS’s leading cloud technology.”

The partnership also encompasses a program called the AWS Generative AI Innovation Center that the cloud giant debuted earlier this year. Through the program, AWS will assign teams of machine learning experts to help customers build generative AI applications. Parent company Amazon, meanwhile, will give its engineers access to Anthropic models for use in product development projects.

“The agreement should further bolster AMZN’s offerings both at the Bedrock and application layer,” TD Cowen analyst John Blackledge wrote in a note to clients today. “Additionally, we view the deal as demonstrative of AWS’ hardware capabilities, namely for companies looking to utilize LLMs, as Anthropic now plans to run a majority of workloads on AWS while utilizing their Trainium and Inferentia chips for building, training, and deployment of future LLM versions.”

Today’s investment comes six months after word emerged that Anthropic was seeking to raise up to $5 billion in funding by 2027. According to internal documents obtained by TechCrunch, the company was planning to use the capital to build a sophisticated foundation model called Claude Next. It’s expected to be 10 times more powerful than the most advanced AI system on the market today.

According to the leaked documents, Claude Next will require computing infrastructure with performance of 10 septillion FLOPs, or floating point operations per second. One septillion is one followed by 24 zeros. That’s several orders of magnitude larger than a quintillion, the number in which the floating point performance of today’s fastest supercomputers is measured.

It’s unclear whether Anthropic will require 10 septillion FLOPs of computing capacity to train Claude Next or to support users’ inference workloads. Either way, building infrastructure capable of such speeds is likely to be a costly endeavor. Amazon’s newly announced investment should make it easier for Anthropic to reach its AI development goals.

Image: Anthropic