Microsoft will let companies run and tune its AI inside their own cloud

A new license puts copies of Microsoft's own AI models in a customer's Azure account, where the company decides who uses them, how they change and when they are upgraded.

Watercolor shapes in pink, blue and orange behind the words Licensed Weights.
Image: Microsoft

The short version

  • Microsoft will let companies put copies of its own AI models in their own Azure cloud accounts and adapt them with company data.
  • The customer controls access to the model and decides when it changes or is upgraded. Microsoft trains on that data only with the customer's express agreement.
  • Early users report big savings on narrow jobs. AT&T says a tuned model handled standard network tasks at 1/50th of the cost of a leading OpenAI model.

What Microsoft announced

At its Ignite conference on 17 November, Microsoft introduced MAI Licensed Weights. MAI is the family of AI models Microsoft builds itself.

A model's weights are the settings it learned in training. In practice, the weights are the model. Licensing them means your company gets its own copy, running in your own Azure environment.

Microsoft frames this around control. Your data and your changes stay yours. You decide who has access. You also decide when to move to a newer version, instead of having the provider switch it for you.

Microsoft says the offer is coming soon. For now, you ask your Microsoft account team.

What is actually new

The usual way to use AI is to send each request to a provider's model and pay per use. The provider runs the model and changes it on its own schedule.

Here the company holds the copy. That brings three practical changes. A model tuned on your own records can learn your terms and your way of working. Costs can fall, because a smaller tuned model may do a narrow job as well as a large general one. And nothing changes under you without your say.

Microsoft gave three customer examples:

  • AT&T tuned MAI models to analyze 5G networks. On standard tasks, the tuned model beat a leading OpenAI model at 1/50th of the cost. On harder tasks it matched top models at a much lower cost.
  • Woven by Toyota trained a model on its past fixes to code that broke its coding rules. It produced ready-to-merge fixes in 44% of the cases where its existing top models had needed a person to step in. It did so at under a tenth of the cost.
  • Alaska Airlines used a tuned model to sort aircraft maintenance reports. Accuracy rose from 81% to 91% after tuning. Qualified staff still make the final decisions.

Microsoft notes that these results come from each customer's own tests, and that savings vary by task.

What it means for companies in Indonesia

Banks and insurers have to show regulators and auditors where their data goes. A model that lives in your own cloud account is easier to explain than one you rent request by request.

The examples share a pattern. Each company picked one narrow job that repeats, and taught the model with years of its own past decisions. None of them asked a general model to do everything.

Most large Indonesian firms have that kind of record already: resolved complaints, approved claims, closed tickets, reviewed contracts. That history is what makes tuning pay off.

The catch is effort. Someone has to choose the task, clean the examples and measure the result. A licensed model is a project, and it needs an owner.

What to do

  • Pick a narrow, high-volume task where you already have years of correct past decisions.
  • Measure what that task costs and how accurate it is today, so you can compare.
  • Ask your Microsoft account team about timing, since the offer is not yet generally available.
  • Decide who in your company owns upgrades, because you now choose when to change models.
  • Keep a person signing off on decisions that matter, as Alaska Airlines does.

Sources

  1. 01 · Microsoft AI · 2026-11-17

Keep reading

  1. Research

    AI assistants do better with fewer notes they check, Anthropic finds

  2. Report

    Two AI models did the same coding work, but one cost 32x more

  3. Release

    Databricks' AI can learn your team's way of doing monthly reports