Microsoft chief executive Satya Nadella’s decision to throw his company’s lot in with OpenAI in 2019 has turned into one of the tech industry’s most successful partnerships, giving the software company a huge head-start in the booming market for generative artificial intelligence.
His closeness with OpenAI chief Sam Altman, underlined when Nadella turned up on stage at the AI start-up’s developer day earlier this month, has also highlighted how much he has personally riding on the alliance.
Beneath the camaraderie, a hard business logic has shaped the relationship. For Microsoft, OpenAI’s language models have become a key ingredient in the company’s latest AI-infused services, while the AI company needs Microsoft’s cloud platform — and money — to support its giant AI models. Speaking in a Financial Times interview this year, Nadella admitted the two companies had developed a “mutual dependence”.
That made last week’s shock sacking of Altman a potential threat to Microsoft’s core AI strategy. He was reinstated five days later, but the episode has highlighted a vulnerability and brought an end to the halo effect from the partnership that has benefited both companies this year. That, in turn, could make it harder for Microsoft to satisfy Wall Street’s hopes that its lead in generative AI will quickly translate into higher revenues, according to some analysts.
After months of development, the software company has only recently increased efforts to sell its new generative AI services, said Jason Wong, an analyst at IT research firm Gartner. “It was pedal to the metal,” he said after the company’s annual Ignite conference early last week. But with the debacle at OpenAI, “that momentum has been broken”.
Many customers are now likely to hesitate before taking the plunge, said Wong: “They will say, ‘are we just investing in this because of the hype?’. This is a good pause for the business community, to take stock of where AI is going.”
Nadella’s handling of the crisis has also appeared to lessen the damage, drawing plaudits from close watchers of the company. He moved quickly to reassure customers that Microsoft’s access to OpenAI’s technology was secure. And while Microsoft has committed to invest $13bn in OpenAI, some of that money won’t be paid until OpenAI meets future commitments, according to one person familiar with the relationship.
Earlier this week, when it seemed that the position was irreparable, Nadella announced that Microsoft would hire Altman and Greg Brockman, the former OpenAI chair who resigned last week. With many employees threatening to jump ship this week if Altman wasn’t restored, the move appeared to leave Microsoft in a position to take on much of OpenAI’s brainpower, without having to pay a premium for the company or its technology.
Trying to rebuild OpenAI inside Microsoft, however, would have presented challenges of its own. Absorbing the company’s staff would have saddled it with an expensive — and culturally very different — research unit, potentially a less appealing alternative than the current arrangement where Microsoft gets many of the benefits without owning the technology outright.
It might also have invited regulatory scrutiny. Earlier this week, when it seemed that much of OpenAI’s staff might move to Microsoft, one Google executive, who declined to be named, said that regulators should be concerned about the possibility. If the former OpenAI staff were able to replicate the company’s AI models inside Microsoft, it would have left the software company’s Azure cloud platform with exclusive access to the technology, this person said.
By contrast, Altman’s return to OpenAI, after Microsoft had offered to hire him, should cement the relationship between the two companies, said Barry Briggs, a former Microsoft executive.
“When Sam didn’t have a home, Microsoft gave him one without hesitation — and when the whole company didn’t have a home, Microsoft gave them one,” said Briggs, who is now an analyst at Directions on Microsoft.
“It’s typical classy Nadella,” added Brent Thill, an analyst at Jefferies, of the chief executive’s handling of the crisis. “I don’t think much has changed. There was too much at stake for everyone not to get along.”
However, the episode has highlighted problems that drew little attention before. For instance, it exposed flaws in OpenAI’s governance arrangements — where a non-profit board oversees a for-profit start-up — that have not been fixed by the hasty pact that saw Altman return to the company. Microsoft is hoping for governance changes that will make OpenAI more stable in future, but it has no direct say in how the company is run.
Microsoft had also already been moving towards becoming less dependent on OpenAI. As customers have started to ask for a choice of AI technology, it has joined Amazon Web Services and Google in offering models from other companies, not just OpenAI.
Yet the tight integration between Microsoft’s cloud platform and OpenAI’s models has been one of Microsoft’s main selling points, making any dilution of the relationship a negative. Optimising OpenAI’s models to run on Microsoft’s hardware has helped to drive down costs and enabled OpenAI to cut prices faster than its rivals.
Customers are more likely to choose to tap into OpenAI’s models through Microsoft’s Azure cloud than go direct, preferring the assurance that comes with buying a service from a bigger technology company with stronger tech foundations, said Briggs.
The software company’s focus on integrating all its AI technologies had led to “very, very high leverage of the one model”, Nadella said on Microsoft’s most recent earnings call.
Microsoft’s business dependence on OpenAI is also starting to fade as the generative AI market evolves. Much of the value in Microsoft’s Copilot services — the intelligent assistants built into its other software products — comes from its own software, rather than OpenAI’s language models, according to Wong. As a result, customers are less concerned about which models are running in the background of services like these, he added.
As generative AI services become more specialised, with language models that are geared to specific industries or tasks, Microsoft’s in-house research efforts are also likely to play a more important role.
Microsoft Research lost out to OpenAI and Google in the race to build large language models, but it is well-placed to develop the next generation of SLMs, or small language models, said Briggs. This week it announced the release of Orca 2, a pair of language models that it claimed could outperform rival systems five to 10 times the size.
Developments like these could reduce Microsoft’s dependence on OpenAI in the next phase of generative AI and leave Nadella with more options. None, however, are likely to repeat the sort of competitive leap that came from his early bet on Altman’s AI research group.
Read the full article here