For years, the Microsoft and OpenAI relationship looked like the cleanest bet in enterprise technology. Microsoft invested, OpenAI built the models, and Microsoft sold them through its cloud and software products. The two companies were so tightly bound that Microsoft was actually contractually barred from building its own advanced AI models while the deal was in force.
That constraint is now gone. The April 27 contract amendment formally ended Microsoft's exclusive licence to OpenAI's models. OpenAI can now sell to Amazon, Google, or anyone else. Microsoft kept its licence to OpenAI's technology through 2032 and retained a 27 percent stake worth roughly $135 billion. But the assumption that Microsoft would only ever need one AI supplier has been dropped from the contract entirely.
Within weeks of signing that amendment, Reuters reported that Microsoft has been quietly approaching AI startups about acquisitions and strategic deals. The goal, according to people familiar with the conversations, is to build the option to operate without OpenAI if it ever needs to.
The clearest example of how this is playing out is Cursor. The startup makes a tool that helps developers write code faster, and it grew to $2 billion in annualised revenue in about three years, which is an unusual pace by any measure. Microsoft considered buying it. Then it did not. The internal verdict was that owning GitHub Copilot, its existing developer tool, and acquiring Cursor simultaneously would draw regulatory attention Microsoft did not want. SpaceX then stepped in and secured the right to buy Cursor for $60 billion, with a $10 billion payment owed if the deal falls apart by year end.
The startup Microsoft is actively talking to now is Inception, based in Palo Alto and founded by Stanford professor Stefano Ermon. Inception builds AI models using a method called diffusion, which is the same technique behind image generators like Midjourney and OpenAI's own image tools. Applying diffusion to text is technically harder, and Ermon's team spent years getting it to work. The practical result is that their models generate answers by refining an entire output in parallel rather than producing one word at a time. Inception claims this makes their models up to ten times faster and ten times cheaper to run than conventional models. Microsoft's venture fund participated in Inception's $50 million funding round last November, alongside Nvidia's investment arm and Databricks. Talks about a larger deal are reportedly now underway.
The reason Microsoft is interested in these two startups is not about raw capability. OpenAI still has the more powerful general-purpose models. What Cursor and Inception both represent is the layer underneath: the tools developers use every day, and a different architectural approach to building models that could eventually be cheaper to operate at scale. Microsoft is trying to own more of that layer before someone else does.
Running parallel to the acquisition search is Microsoft's own internal model-building effort. The MAI Superintelligence team, formed in November under Mustafa Suleyman, the co-founder of DeepMind who joined Microsoft in 2024, shipped its first three foundation models in April: tools for transcription, voice generation, and image generation. A full-scale general-purpose language model is targeted for 2027. The original Microsoft-OpenAI contract had actually capped how large a model Microsoft could train on its own. That cap is now gone.
For business operators, the practical implication is not complicated. Microsoft products, from Office Copilot to Azure AI services, are going to be powered by a mix of models over the coming years: some from OpenAI, some built in-house, and potentially some from startups like Inception. The tools that sit in front of workers every day are increasingly the output of a procurement strategy, not a single supplier relationship. That gives Microsoft more flexibility, and it means the tools businesses rely on will keep changing as Microsoft rotates the models underneath.
The thing worth watching is whether Microsoft actually closes an acquisition before 2025 ends or lets the startup pipeline sit as an option. After the Cursor episode, where hesitation cost Microsoft the asset entirely, the pressure to move faster has gone up considerably.