TheChek News Flow

The Takeaway

As Amazon weighs an equity investment of tens of billions of dollars in OpenAI, the companies are also discussing a commercial agreement that could require OpenAI to dedicate its own researchers and engineers to developing customized models for powering Amazon’s own AI products, according to a person involved in the talks and another person who was briefed about it.

Amazon could use customized versions of OpenAI’s models to bolster Amazon AI products such as the Alexa voice assistant, one of these people said. Such an arrangement could require both companies to tweak OpenAI models so they respond to customers the way Amazon wants, this person said.

A spokesperson for OpenAI said, “We are focused on our strong existing compute partnership with Amazon.” An Amazon spokesperson declined to comment.

While such a deal could boost revenue at OpenAI, it could also divert the company’s resources at a time when it faces growing competition from Google in consumer chatbots and Anthropic in developing AI for application developers and businesses.

It isn’t clear how much of a discount Amazon would get from OpenAI if it uses the ChatGPT maker’s customized models.

The deal discussions show the lengths to which OpenAI may go in the effort to raise enough capital so it can meet its spending commitments to cloud providers. Those commitments total more than $600 billion between now and the early 2030s. Nvidia, Amazon and Microsoft together could contribute more than $60 billion in new funding to OpenAI, The Information has reported, showing how the company’s cloud providers and chip partners are helping it make such payments.

Such a deal might also help Amazon’s efforts to offer consumer and business AI software that stands out in a crowded market. Anthropic’s technology currently helps power a variety of Amazon AI consumer and business products, including shopping assistant Rufus, an AI coding agent, and a tool for searching corporate data.

But Amazon employees have faced restrictions on customizing models from Anthropic, people who have worked on the products said. Meanwhile, Amazon’s effort to create its own AI models, in part for use in its own products, haven’t lived up to its expectations, The Information has reported.

The pending Amazon-OpenAI megadeal shows how the two companies are increasingly mirroring the kind of circular arrangement Microsoft and OpenAI previously forged at the start of the AI boom, in 2022.

Last fall, for instance, Amazon and OpenAI unveiled a deal for OpenAI to rent tens of billions of dollars’ worth of AI servers from Amazon Web Services, similar to an even bigger server rental agreement OpenAI had struck with Microsoft. Then, last week, as OpenAI sought $100 billion in new capital to fund its expensive AI development, Amazon was in talks to put in more than $20 billion into the round, though that figure could change.

As part of the ongoing talks, the company has also talked to OpenAI about renting Amazon’s in-house Trainium AI chips. (OpenAI currently rents Nvidia chips from AWS.) If Amazon gains OpenAI as a customer of the Trainium chips, that could help the cloud giant improve its servers for other cloud customers and potentially lower the amount of money it spends on Nvidia chips.

The two sides have also discussed a potential commerce partnership, The Information previously reported. Amazon has locked down its retail site against OpenAI and other outside firms’ shopping tools, but Amazon CEO Andy Jassy recently said the company is open to AI commerce partnerships if the terms are agreeable.

At the start of the boom, Amazon struck a close relationship with OpenAI’s archrival, Anthropic, but didn’t get many of the special privileges Microsoft got as an early OpenAI benefactor.

For example, Microsoft’s agreement with OpenAI allows it to use OpenAI models for free and to heavily customize such models for its own purposes, in addition to accessing the firm’s intellectual property. (That hasn’t stopped Microsoft leaders from dueling with their OpenAI counterparts about the details of the arrangement along the way.)

Anthropic is generally more restrictive about letting customers customize its models, a process known as fine-tuning or post-training, than OpenAI is. For instance, Amazon employees cannot post-train Anthropic models beyond what other Anthropic customers have access to, according to two people with knowledge of the matter. On the plus side, Amazon employees have been able to use Anthropic models to create less powerful models through a process known as distillation, the people said.

Amazon not only depends on Anthropic for its own products, but also relies on Anthropic models to power a key cloud service, Bedrock, that enables businesses to use AI for developing and running applications. While Amazon sells hundreds of AI models through Bedrock, as of a year ago, much of the revenue from that business was coming from sales of Anthropic models, according to a person who worked on the product.