Where Microsoft and OpenAI stand now

One is chasing models, while the other feels its needs can’t be met by just one partner.

The Microsoft-OpenAI partnership has been on shaky ground for months now, as the two companies have started to compete across more areas and formed new alliances with the other’s rivals. But the companies recently clarified a new phase of their agreement, as OpenAI continued its drift toward a more traditional for-profit structure.

Freed from certain strictures of their previous deal, per Fortune, Microsoft then debuted a new superintelligence team that will allow it to build its own more sophisticated models in a potential bid for AI self-sufficiency. Microsoft released its first in-house models in August.

Meanwhile, OpenAI has woven an elaborate web of deals to feed its insatiable need for more infrastructure to undergird its AI ambitions, including agreements with Microsoft rivals Oracle, Google, and Amazon. OpenAI now plans to spend more on Oracle and others by 2030 than it will with Microsoft, The Information reported.

Analysts said it was inevitable that the two companies would outgrow the alliance that has defined the generative AI revolution.

“It’s been pretty clear that these two are at least going to be frenemies,” Larry Dignan, editor in chief at Constellation Research, told us. “Today, it makes sense to be commingled and have this relationship, but they’re going to gradually separate because I think they’re going to gradually compete.”

Playing catch-up: Under the terms of the new OpenAI deal, Microsoft has the rights to the AI lab’s models and product intellectual property through 2032, and research until 2030.

One unusual feature of the deal remains: If OpenAI declares that it’s achieved artificial general intelligence (AGI) before that deadline—a claim that now must be verified by independent experts—then those research rights will expire early, but model rights persist “post-AGI, with appropriate safety guardrails.”

Because of Microsoft’s reliance on OpenAI, its in-house models lag those of rivals like Google and Amazon, according to Jason Wong, distinguished VP analyst at Gartner.

“This [deadline] gives [Microsoft] a runway for putting out their own models,” Wong said. “As one of the larger vendors, they are behind in doing that…So number one is assuring access for the rest of this decade, a timeline that they can apply their learnings from their partnership to build their own in-house models.”

New terms: Microsoft’s stake in OpenAI’s for-profit entity has been diluted from 32.5% to 27% and it no longer has right of first refusal to be OpenAI’s computing provider, but OpenAI is on the hook for $250 billion more in Azure cloud services.

OpenAI’s APIs are still exclusive to Azure among cloud platforms, until AGI, but Microsoft no longer has rights to any consumer hardware that OpenAI develops or exclusivity on non-API products, like the applications that OpenAI is increasingly interested in developing.

“You can see [OpenAI] trying to break out from just being the model provider and into creating applications,” Wong said. “They are moving down the route of creating applications. They hired their own CEO for applications. They obviously have intentions on the enterprise and looking at specific enterprise workloads. They are looking at…different types of agents that they can potentially sell as applications or agent services.”

It’s that foray into enterprise applications, more than consumer AI or hardware ambitions, where Microsoft and OpenAI are most likely to butt heads, according to Dignan.

“They’re ultimately going to split on some level…It’s basically going to be who owns the customer relationship,” Dignan said. “Obviously, OpenAI and Sam Altman want it, but Microsoft isn’t going to give that up. So I expect them to diverge.”