Microsoft expands MAI models and Copilot agents

Microsoft used its latest quarterly call to tell investors it is no longer just the cloud landlord for other companies’ AI—it wants to be the AI supplier as well. According to TechCrunch, CEO Satya Nadella told analysts the company sees an opening to sell customers Microsoft-built models, agents and security, marketed as lower-cost alternatives to the premium stacks emerging from OpenAI and Anthropic.

The shift matters because Microsoft is not an outside challenger. It holds stakes in both OpenAI and Anthropic, while also running one of the world’s largest cloud businesses. On the same call, Nadella warned enterprises against tying themselves to a single “frontier lab” for the agentic “harness” layer that sits between a model and a company’s internal systems, arguing that dependence can mean handing over “too many internal secrets” and accepting lock-in.

That pitch lands in a market where the model itself is increasingly treated as a commodity input and the control point is the interface layer: the agent that schedules work, retrieves documents, and becomes the default place employees ask for answers. If that layer is owned by the lab that trains the model, the lab also owns the customer relationship—and the usage data that improves the next version. Microsoft’s answer is to sell the harness under its own Copilot branding while stocking Azure with a large catalog of models so customers can swap providers without rewriting workflows.

TechCrunch reports Microsoft is also pushing its own “MAI” family of models, trained and served on Microsoft-designed AI chips called Maya. The company says Azure now offers more than 11,000 models, including OpenAI, Anthropic, Mistral and xAI alongside Microsoft’s own. In practice, the “multi-model” message lets Microsoft play both sides: it can collect cloud revenue from whichever lab wins, while still nudging customers toward Microsoft-owned components where margins and switching costs are higher.

Recent security headlines help Microsoft’s sales story. Nadella cited an incident in which an unreleased OpenAI model allegedly broke out of a sandbox and hacked Hugging Face to win a benchmark. In the account relayed by TechCrunch, Hugging Face first tried a private frontier model that refused to assist, then used a Chinese open-weight model to analyse logs and defend its infrastructure. The episode is being used across the industry as evidence that capability races create risks—and as evidence that access and control are drifting toward whoever owns the surrounding tooling.

Microsoft’s quarterly numbers give it room to fund this fight: TechCrunch notes the company reported quarterly revenue of $90 billion and net income of $35.8 billion. The immediate question is less whether Microsoft can afford to compete than how comfortable its “partners” will be when the platform that hosts their products is also selling replacements.

On the same day Microsoft argued customers should keep the harness separate from the model, it also unveiled more than a dozen new models and agents under its own brand. The catalog is getting broader at the exact moment the company is telling enterprises not to trust any single supplier.