Microsoft is looking to become more self-sufficient in the artificial-intelligence sphere.
The tech giant is pursuing new research avenues that the company hopes will put its AI models on par with rivals, after inking a reworked deal with ChatGPT parent OpenAI in 2025.
Microsoft’s prescient 2019 initial investment in OpenAI came with an agreement that tied the two companies together in more than just financial ways. Microsoft’s Azure cloud unit, which was still getting off the ground, would become OpenAI’s sole cloud provider. On the flip-side, Microsoft would not compete with OpenAI to build the ultimate goal of AI researchers: machines that can perform knowledge work better than any person, also called superintelligence.
But as ChatGPT boomed in 2023 and 2024, both OpenAI and Microsoft began re-assessing the marriage: neither wanted to be tied to a single vendor for something so important. As part of an OpenAI restructuring last fall, the agreement was loosened so that OpenAI could use other cloud providers, and Microsoft could pursue its own version of superintelligence.
At the time, Microsoft still owned about 27% of OpenAI. That stake may have been diluted this week after OpenAI’s $122 billion fundraising round closed.
Microsoft has begun a program to produce “frontier” models to compete with ChatGPT, Anthropic’s Claude, Google’s Gemini, and some Chinese models. Just after the OpenAI semi-divorce, Microsoft AI CEO Mustafa Suleyman announced a new superintelligence unit at Microsoft in November.
This was the beginning of Microsoft’s effort to own more of its AI technology stack. “We have to develop our own foundation models which are at the absolute frontier,” Suleyman said in a February interview with the Financial Times. “That’s our sort of true self-sufficiency mission.”
OpenAI did not immediately respond to a request for comment.
Every company’s take on what superintelligence means is different, and Microsoft’s is on-brand: Superintelligence as a Microsoft enterprise product that works within the strict bounds its customers require. “We think of it as systems that are problem-oriented and tend towards the domain specific,” said Suleyman in a November blog post. “Not an unbounded and unlimited entity with high degrees of autonomy – but AI that is carefully calibrated, contextualized, within limits.”
In March, CEO Satya Nadella reorganized the effort. Suleyman handed over responsibilities for AI services, such as the expensive Microsoft 365 Copilot AI assistant, so that he could be laser-focused on new frontier models.
But in January, a crucial internal AI resource battle bubbled to the surface during the company’s second-quarter’s earnings call. Investors were disappointed with Azure sales growth and in explaining the miss, chief financial officer Amy Hood said that Azure sales could have beaten expectations, but that Microsoft decided to reserve data center capacity for its own services and research. The $160 billion in Microsoft capital expenditures in the past six quarters couldn’t satisfy all internal and external computing demand at the same time.
Investors were not happy with that trade-off, and the stock declined 10% the next day. But the tension between renting out AI servers and using them internally has had effects on Suleyman’s project.
This week, while releasing new AI models for internal uses, such as transcriptions, he described them to the Financial Times as “mid-class” and not in competition with the best frontier models. He cited the internal compute shortage as the limiting factor, which he predicted would clear up this year for his team.
Microsoft is on a new AI journey. It may take a while, but never count the company out.
Write to Adam Levine at adam.levine@barrons.com