Microsoft’s MAI Cost Cut Reveals the Switchboard

Microsoft says it cut image-generation costs in PowerPoint by about 85% after replacing OpenAI with MAI. Two days earlier, it expanded its outside-model roster with Mistral.

In PowerPoint and Bing, Microsoft installed its own image models. Its Mistral agreement puts another external supplier into Foundry, Copilot Studio, and Azure Local, alongside plans to build European data centers together. One move narrows the supplier list inside Microsoft’s products; the other widens it across Microsoft’s platforms.

Microsoft is making the endpoint model-agnostic. Users remain in PowerPoint, Bing, Excel, Outlook, Copilot Studio, or Azure while product teams change the model underneath. Microsoft can draw on frontier, specialized, small, and open models without handing the workflow to any one supplier.

Microsoft turned dependency into a unit-cost decision

Microsoft turned to OpenAI to put frontier AI into established products before it had equivalent models of its own. In 2023, Microsoft launched the new Bing with an OpenAI large language model and its Prometheus system. Microsoft supplied the search engine, interface, cloud infrastructure, and distribution, while OpenAI supplied the capability that made the product newly useful.

Microsoft used the partnership to shorten the path from model research to a consumer product. As those features expanded, however, every routine call exposed Microsoft to the economics of a single outside supplier. By September 2023, reporting described the company developing efficient models that could approximate more capable OpenAI systems to reduce the cost of features such as Bing Chat.

Microsoft built its substitution capacity in stages. It introduced MAI-Image-1, its first in-house text-to-image model, in October 2025. In March 2026, MAI-Image-2 ranked third on Arena AI’s text-to-image leaderboard, behind models from Google and OpenAI. The following month, Microsoft released an efficient variant that it said delivered production-ready quality at roughly half the cost.

With those releases, Microsoft tested whether MAI could approach frontier quality, lower costs, and carry production traffic. It answered by assigning PowerPoint and Bing workloads to its own models.

Third place was enough to win the workload

Microsoft had a viable substitute once MAI-Image-2 reached third place on Arena AI’s text-to-image leaderboard. Google and OpenAI still ranked ahead of it in March.

Arena AI ranks relative quality. Microsoft pays for every inference, so its product teams must decide whether capability above a task’s quality threshold creates enough value to justify the added cost. After a model clears that threshold, teams can optimize for cost per useful task.

The PowerPoint team can accept a modest capability gap when the reported cost difference is so large. Google or OpenAI can remain better-ranked while Microsoft captures the saving on each qualifying call as the bulk buyer of image-generation inference.

Microsoft can spread that per-call saving across large product surfaces. Bing had crossed 100 million daily active users during the 2023 launch period. Separate July reporting said the company was also beginning to replace OpenAI and Anthropic models with MAI in products including Excel and Outlook.

Microsoft combines integration with supplier choice

By May 2024, Microsoft had made Azure AI Studio broadly available with support for OpenAI’s GPT-4o and its own Phi-3 family. The company placed model choice inside its development, integration, and governance environment.

Through the Mistral agreement, Microsoft extended that catalogue across Foundry, Copilot Studio, Azure Local, and planned European data centers. Customers bring different quality requirements, deployment constraints, regions, and price points, giving Microsoft a reason to keep outside models central even as it develops MAI.

Product teams assign workloads, procurement groups approve suppliers, and developers select models in Foundry. Together, those teams orchestrate the portfolio before software routes requests in real time.

Hyperscalers sell choice because choice protects them

Microsoft’s posture fits a broader reordering among cloud owners. Amazon planned to offer OpenAI’s gpt-oss models through Bedrock and SageMaker, the first time OpenAI models would be available to AWS customers. Amazon treated OpenAI as another option inside its distribution system.

Microsoft extends supplier choice below the model layer. It plans to deploy AMD’s Helios rack-scale platform on Azure for frontier-model inference, adding another hardware supplier beneath its model catalogue. With OpenAI, MAI, and Mistral at the model layer and AMD among the hardware suppliers beneath it, Microsoft can negotiate across two layers of the inference stack.

Customers can still demand the strongest available system or name a particular vendor. A broad catalogue lets Microsoft and Amazon meet either demand while retaining the infrastructure sale.

Supplier choice cannot erase the capacity bill

Barclays projected that inference capital expenditure would surpass training within two years and reach $208.2 billion in 2026. Documents reviewed by The Wall Street Journal showed OpenAI and Anthropic reporting inference costs exceeding half of revenue.

Model developers incur large training bills, and every product use creates a recurring inference cost. For Microsoft, a single user action can trigger repeated model calls as copilots expand into agents and AI features enter ordinary software. Calling the most expensive model when a cheaper one would clear the task’s quality threshold sacrifices margin across the installed base.

Microsoft can reserve higher-cost capability for workloads that require it, assign cheaper specialized models elsewhere, and bargain across model and hardware suppliers. It must still finance racks, chips, cooling equipment, fiber, substations, power contracts, and buildings before their utilization is known.

Cloud owners have less room for error as the model layer becomes more flexible. A study estimated that off-balance-sheet debt at Alphabet, Microsoft, Amazon, Meta, and Oracle had grown roughly eightfold since 2022 to $1.65 trillion, above an estimated $1.35 trillion in balance-sheet debt. Microsoft and its peers can bargain across suppliers, but they still carry the demand risk attached to the capacity they commission.

Microsoft still has to fill contracted megawatts with useful inference. A cheaper MAI call improves PowerPoint’s economics and raises the useful work each committed server and megawatt can deliver.

PowerPoint’s image button now leads to a switchboard Microsoft owns. Microsoft says MAI runs that workload for about 85% less; OpenAI can still supply frontier capability, Mistral adds another circuit, and Microsoft keeps the meter.