What if Microsoft let OpenAI go free?

So far it has been a wildly successful partnership. Funding from Microsoft has helped OpenAI build bigger and better large language models (LLMs). That technology, in turn, has been incorporated into Microsoft’s various software products. The ties between the two have allowed Azure to chip away at the lead of AWS, Amazon’s cloud-computing division, notes Brent Thill of Jefferies, an investment bank (see chart). Analysts on average reckon Azure’s revenue grew by 30%, year on year, in the quarter from July to September, compared with 19% for AWS; the rivals both report their results this week. Eric Boyd, head of Microsoft’s AI platform, says the “incredibly fruitful” partnership with OpenAI has helped lure 60,000 customers to Azure AI.

Yet like Michelangelo with the Medicis, OpenAI sometimes chafes at the ties to its wealthy benefactor. Some of its board members and other investors have told The Economist they believe that Microsoft should loosen its grip. In particular, they have their eyes on the half of the cloud-computing market still controlled by AWS. Gaining access to that would increase OpenAI’s already dominant position in the provision of LLms, lifting revenues that are already expected to be upwards of $3.5bn this year.

Microsoft declines to discuss details of its contract with OpenAI. Given Azure’s long-standing rivalry with AWS, it is a fair bet the Redmond-based company is loth to lend its Seattle-based neighbour a helping hand. For Microsoft, however, ending its exclusive patronage of OpenAI may not be as crazy as it sounds.

Those championing more commercial freedom for OpenAI argue that although Microsoft might resent sharing models with AWS, in the long run its equity stake in OpenAI would mean it would benefit from the model-maker gaining broader market access. Antitrust concerns also bolster the case for giving OpenAI greater independence. America’s Federal Trade Commission and Britain’s Competition and Markets Authority have opened inquiries into the relationship between Microsoft and OpenAI. Mr Thill of Jefferies says more openness would be in Microsoft’s long-term interest. “It’s like a kid going to college,” he says. “It’s painful. But the right thing is to let [OpenAI] out into the world.”

Microsoft has already begun reducing its dependence on OpenAI. Satya Nadella, the tech giant’s boss, was reportedly shocked when Sam Altman, his counterpart at OpenAI, was briefly ousted last November, before being quickly reinstated. Since then Microsoft has been hedging its bets, including by adding LLMs from Mistral, a French AI firm, and other model-makersto its line-up and hiring nearly all the staff of Inflection, an OpenAI rival, including its boss, Mustafa Suleyman (who sits on the board of The Economist’s parent company).

Microsoft and OpenAI are in the process of renegotiating the terms of their relationship as the latter changes its corporate structure from a non-profit to a profit-making entity. There may also be a looming sunset clause. OpenAI is believed to have the right to dissolve its commercial ties with Microsoft if its models reach a level of superhuman capability called artificial general intelligence. What that means in practice is subjective, but some AI enthusiasts argue it could be only a few years away.

Amazon, for its part, would be delighted to gain access to OpenAI’s models. “I would love for OpenAI to run on AWS,” says Matt Garman, AWS’s newish boss (though he will not be drawn on whether formal discussions are taking place). Pradeep Sanyal, an AI consultant and former AWS executive, says that even though Amazon’s cloud business is still the biggest, it is “third in mindshare” when it comes to generative AI, behind Azure and Google Cloud. Amazon has neither a big software business where it can demonstrate its AI capabilities nor an LLM whizzy enough to compete with those of OpenAI or Google.

AWS offers customers a wide variety of models, ranging from those of Anthropic (in which Amazon is a big investor) to Meta’s Llama family of open models (which Mr Garman says are very popular among AI startups). Adding those of OpenAI would no doubt help it attract customers. But Mr Garman does not believe any one model will dominate completely; OpenAI, Anthropic and other model-makers are currently “leap-frogging each other”, he notes. On October 22nd Anthropic launched an experimental version of its Claude 3.5 Sonnet model that can use a computer the way a human does, including virtually operating the keyboard and mouse.

Amazon’s more open approach may yet prevail. Eric Sheridan of Goldman Sachs, a bank, thinks it will take years, not months, to determine which of the different cloud-service providers emerges as the definitive winner in generative AI. He points out, though, that the long-running trend in cloud computing is away from exclusivity towards more open relationships. Companies increasingly use more than one cloud-service provider and may benefit from using different LLMs for different functions. Like the Medicis, Microsoft may well go down in history for having spotted creative genius early on. But its hold over OpenAI may not last forever.