# MAI Licensed Weights, Microsoft's models in your Azure

> Microsoft offers to customize a frontier MAI model and run it in the customer's Azure without publishing the weights. AT&amp;T brought it to a cost 50 times lower.
> By José Maldonado · Artificial Intelligence · 2026-11-19

MAI Licensed Weights is Microsoft's answer to a question data teams have been asking quietly for two years: if the model that runs my business isn't mine, what happens when it changes price, owner or criteria. The formula that [Microsoft presented at Ignite](https://microsoft.ai/news/introducing-mai-licensed-weights/) on November 17, 2026 consists of taking a frontier model from the MAI family, customizing it with proprietary data and running it inside Azure, in the customer's environment, with control over who accesses it and without publishing the weights.

The announcement is better understood as a move in the fight over model ownership than as a new product. In 2026 the discussion no longer separates open from closed with the clarity of two years ago: half the industry publishes downloadable weights and the other half rents access per token, with the price per million that Anthropic set for [its two frontier models](https://observatorioblockchain.com/en/ai/claude-fable-5-2-mythos-5-2-anthopic-ipo/). Microsoft adds a third path, with the weights on its side of the contract and execution on the customer's side.

The fine print starts with availability. The offer reaches Frontier Company program customers first, and none of the pages published these days includes a rate, so the price is negotiated with the Microsoft sales rep and is not read off a list.

## From open weights to licensed weights

A weight is a number the model learned during its training, and the set of all of them is what makes it respond in a particular way. Downloading weights and licensing them are barely alike. The first allows you to run the model and, depending on the license, fine-tune it. The second is a contract that says what you can do with the result, how much you pay and what happens if the provider changes its mind about the product it rented you.

The difference is measured against a written definition. The [OSI's open source AI definition](https://opensource.org/ai/open-source-ai-definition) requires that anyone be able to use and study the system, modify it and share it, and have enough data to recreate it. MAI Licensed Weights covers customization and use, not redistribution, and it does not deliver the training data that would be needed to rebuild the model from scratch.

For a bank or an airline, the contractual question is what stays in-house if the provider withdraws that model tomorrow. With downloadable weights a file remains; with a license, whatever the contract says remains. The problem is not technical, it is one of ownership.

## What changes for those who run the model in their own cloud

The announcement text describes the mechanism precisely. The model is adapted with customer data and runs in their Azure environment; the customer decides who accesses it, and the data and adaptations remain theirs. Microsoft would only use them to train its own models when there is an express agreement, drafted in the language of enterprise contracts of recent years.

That access control is worth as much as performance. In a regulated entity, the auditor's question is not what percentage the model scores on a benchmark, but who can invoke it, in which region it runs and what log remains of each call. A rented model answers those questions with the provider's policy; one deployed in the customer's cloud answers them with its own identity and network configuration. If you manage infrastructure for an audited sector, there lies the practical difference between the two purchases.

The other promise is the schedule. The company decides when the model is updated, which on paper prevents someone else's improvement from breaking a production workflow, and in turn forces it to keep several versions alive at once. That freedom has a bill: every version left behind is one someone is still watching.

The [announcement of MAI Licensed Weights at Ignite](https://microsoft.ai/news/microsoft-ignite-2026/) sums up the idea in one sentence: nobody wants to pour their IQ into someone else's model. The same announcement raises a legitimate doubt about the origin of the open weights.

## AT&T and two more customers that back the argument

The first case cited is AT&T, which used MAI models for 5G network operations analytics. On standard tasks, the customized model beat GPT-6.1 Sol at one-fiftieth the cost, according to the results note Microsoft publishes alongside the announcement. In more complex scenarios, the company describes frontier performance at a considerably lower price, without giving firm figures for that comparison.

The other two customers extend the pattern to less glamorous ground. Woven by Toyota fine-tuned MAI Code 2 Flash with its history of MISRA standard corrections and obtained patches ready to merge in 44% of the cases that previously required human intervention, at less than a tenth of the cost, according to Microsoft's [model fine-tuning documentation](https://microsoft.ai/models/microsoft-frontier-tuning/) .

Alaska Airlines went from 81.2% to 91.1% accuracy when classifying aircraft maintenance reports, an improvement the announcement note attributes to fine-tuning the model with the airline's own data.

Microsoft qualifies those numbers in the footnote: they are evaluations done by each customer for their tasks, and the advantages depend on the workload, the configuration and the measurement method. There is no public, independent benchmark comparing the licensed model with the same model served via API: the three results are published by Microsoft, not by an outside lab.

## The price that does not appear in the note

Microsoft does not publish the license price. The announcement page refers to the account team and to Frontier Company agreements, so any figure circulating without a contract behind it is a third-party estimate. The silence is not an exception in this market, but it complicates comparison with the obvious alternative, which is paying per token on the API.

The prices that are published belong to something else. At the same Ignite, MAI Code 2 was announced with 98% of GPT-6.1 Sol's quality at about 40% of the price, and MAI Cyber 2 Flash at less than 2% of the cost per task versus GPT-6 Astra. Those are per-million-token rates for a rented service, and the [flat price Google set for Gemini 4 Hydrogen](https://observatorioblockchain.com/en/ai/gemini-4-hydrogen/) belongs to the same family.

Comparing a license with the API without saying this is comparing a rental with a purchase. A license is negotiated with other variables, such as reserved compute, support and contract years.

The trap of the easy comparison is that: the real cost is not measured only in the model's bill. Hardware, energy and the team that operates it must be added, and those savings coexist with an industry that keeps pouring money into compute capacity, as shown by [TSMC's investment in Anthropic and OpenAI](https://observatorioblockchain.com/en/ai/tsmc-invierte-anthropic-openai-1000-millones/). For a mid-sized company, the useful question is not how much the token drops but how quickly the investment pays back.

## What remains to be seen in the weights license

It remains to be known whether the customer will be able to audit the model it runs. A weights license without access to the training code or the source data leaves the buyer with the provider's word on the system's behavior, and that clashes with the traceability any audit of a high-risk system will demand. The [legal pressure on the big AI providers](https://observatorioblockchain.com/en/ai/demanda-ia-cinco-gigantes/) points in the same direction.

It also remains to be seen what happens when the contract ends. A model customized with proprietary data and run in the customer's cloud looks like an asset, but it only is one if the license allows continued use after the commercial relationship ends. Microsoft has sold control, and the real size of that control will be known in the first contract made public.
