Microsoft reportedly struggling to build its own reasoning models to rival OpenAI

Microsoft may be “all in” with AI, but its efforts to produce homegrown products to feed its ambitions have so far been lagging and its reliance on OpenAI prevails.

However, recent events may augur change. The company has reportedly finally completed training a new family of generative AI models, known as MAI, that performed “nearly as well” as OpenAI and Anthropic models on benchmarks.

The report from The Information said that Microsoft is also training reasoning models that could compete with models from OpenAI, as well as those from DeepSeek and Alibaba. Reasoning models are designed to simulate human cognitive processes such as logical reasoning, decision-making, and problem-solving. A key component is chain of thought, which improves the performance of the models by generating intermediate steps for problem solving to make complex problems more manageable.

OpenAI was not sharing technical information about the chain-of-thought process as agreed, the report noted, which became a source of conflict between the partners.

Technical setbacks and delays

When the news of MAI broke almost a year ago, Microsoft CTO Kevin Scott wrote on LinkedIn, “I’m not sure why this is news, but just to summarize the obvious: we build big supercomputers to train AI models; our partner OpenAI uses these supercomputers to train frontier-defining models; and then we both make these models available in products and services so that lots of people can benefit from them. We rather like this arrangement. We’ve been at it for almost five years now.”

He added, “We also, for years and years and years, have built AI models in [Microsoft Research] and in our product groups.”

However, at the time, speculation had abounded that MAI would be released at the 2024 Microsoft Build conference, and that didn’t come to pass.

The delay, the report said, was caused by “technical setbacks, abrupt changes in strategy, and the departures of top talent who disagreed with [Microsoft AI CEO] Suleyman’s management and technical approach.” And as Microsoft struggled, OpenAI pulled farther ahead with its models. Releasing a preview of o3 and GPT-4.5.

The Information reported Friday, citing unnamed sources, that Microsoft is considering releasing MAI later this year as an application programming interface (API). This, it said, would put Microsoft into direct competition with APIs from rivals and partners alike. It also said, ”the company is already experimenting with swapping out the MAI models … for OpenAI’s models in Microsoft Copilot.”

It noted that Microsoft is testing out other models in Copilot as well, including those from Anthropic, xAI, DeepSeek, and Meta.

A potential rift in the partnership came from the Stargate Project, in which OpenAI is collaborating with Oracle and Nvidia to build and operate AI infrastructure, although OpenAI said during the announcement that the project will build on its existing partnership with Microsoft.

However, Microsoft has in turn altered some of its expansion plans, cancelling planned leases on new data centers, although it hasn’t reduced its expected capital expenditure.

Different priorities

Jason Andersen, vice president and principal analyst, Moor Insights & Strategy, noted that the two companies have different priorities. “OpenAI focuses on delivering the best models in service of the mission of achieving AGI [artificial general intelligence]. There are not many (or any) multi-billion dollar revenue engines with that level of single-mindedness. I believe that Microsoft is dealing with a wide range of AI priorities and that’s precisely why [Microsoft AI CEO] Suleyman was brought in to balance these efforts,” he said. “This range of priorities is also why we have seen Microsoft partner with other LLM [large language model] companies despite having its own products and the OpenAI partnership. Shifting to a multi-model mindset is significant and reinforces the point that right now model pure plays will have an innovation advantage in areas where they focus.”

Justin St-Maurice, technical counselor at Info-Tech Research Group, added that the firms’ technologies are very different as well.

“Instead of operating through a series of functions with predictable inputs and outputs, LLMs accept any arbitrary input. They behave chaotically, with natural system properties,” he observed. “Operating this new type of technology has been OpenAI’s bread and butter since day one, whereas Microsoft has a legacy of thinking, processing, and designing that are fixed in an old paradigm. That said, Microsoft Research has been working on large-scale AI models for years now, so while OpenAI has had a head start on embracing the new paradigm, Microsoft isn’t starting from zero.”

And, he added, “OpenAI’s magic and innovation isn’t just in its underlying LLM; it’s how they decompose, orchestrate, and chain their prompts. While there’s speculation that they chain multiple LLMs in a unique way, their edge comes from a broader set of innovations, including reinforcement learning from human feedback and customer architectures. The secret sauce of every proprietary breakthrough might not be shared with Microsoft, but their deep partnership means they aren’t entirely in the dark either.”

Can Microsoft catch up?

Neither analyst is betting against Microsoft eventually catching up, though.

“As to whether Microsoft will ever catch up is an open question because it will ultimately come down to balancing across priorities,” Andersen said. “This includes innovating in areas where OpenAI is not focused, such as developer tooling, application integration and management, and governance.”   

St-Maurice agreed. “Betting against the turtle in favor of the hare was never a great bet,” he said. “Microsoft will endure, iterate, and optimize in mature labs with world-class engineers. If OpenAI isn’t interested in sharing all of its underlying IP,  the open-source market is providing more than enough output to ensure it doesn’t really matter.”