Microsoft Expands Use of In-House AI Models to Reduce Costs

July 8, 2026 Ai 5-8 min read

Microsoft Expands Use of In-House AI Models to Reduce Costs

For much of the past three years, Microsoft's AI strategy has been inseparable from its relationship with OpenAI. The $13 billion investment, the integration of GPT models into Copilot, the public messaging that positioned Office 365 as powered by best-in-class external AI, all of it built an image of Microsoft as the company that supercharged its products by plugging into the world's most capable AI lab. That image was accurate, and it drove real product improvements. It also came with a financial reality that was becoming increasingly uncomfortable: paying licensing fees to external AI providers at the scale Microsoft operates is an enormous ongoing cost that compounds with every user who generates a Copilot prompt in Excel or Outlook.

Bloomberg reported on July 7, 2026 that Microsoft has quietly begun routing a portion of those user prompts to its own internally developed MAI models rather than to OpenAI or Anthropic. The shift is already running at tens of thousands of prompts per week across Excel and Outlook. The partnerships with OpenAI and Anthropic remain intact, but Microsoft is no longer treating them as the exclusive or even primary layer for every inference request. The company is managing its AI cost structure the same way any large enterprise eventually manages its most expensive dependencies: by building enough internal capability to have genuine options.

Microsoft is expanding the use of its in-house AI models to reduce infrastructure costs and strengthen its AI ecosystem. This article explores the company's strategy, the potential impact on AI development, and what it means for Microsoft's long-term competitiveness in the AI market.

What the MAI Models Actually Are

The MAI family, short for Microsoft AI, was first introduced publicly at Microsoft Build 2026 in May. At that event, Microsoft announced seven in-house models spanning multiple capability categories, with some being variants on shared base architectures. The Build announcement was the public introduction of a model development program that had clearly been running in some form before the conference, given that the models were presented with performance benchmarks and product integration timelines rather than as early research previews.

The flagship model in the family is MAI-Thinking-1, described as Microsoft's first reasoning model. It is built for efficiency and performance simultaneously, and Microsoft's blog post at launch emphasized token cost as explicitly as it emphasized capability. The company highlighted that MAI-Thinking-1 matched Claude Opus 4.6 on coding tasks, a meaningful claim given that Anthropic's model has been one of the leading coding benchmarks. The fact that Microsoft led with the cost efficiency of its flagship rather than positioning it purely as a raw capability leader signals that cost reduction is genuinely the primary motivation rather than a secondary benefit of building proprietary models.

The rest of the MAI family covers additional capability categories that collectively address the full range of AI tasks Microsoft needs to run at scale across its products:

MAI-Code-1-Flash: Microsoft's first proprietary AI coding model, designed for code generation and completion tasks in developer tools including GitHub Copilot
MAI-Image-2.5: Text-to-image generation and image editing, which Microsoft claimed surpassed the Arena score of Google's Nano Banana Pro at launch in May
MAI-Image-2.5 Flash: A faster, lower-cost variant of the image generation model for high-volume generation use cases
MAI Transcribe-1.5: Speech-to-text transcription model intended to power Teams and other Microsoft products requiring real-time or batch audio processing
MAI-Voice-2: Speech generation across 15 languages for voice interface applications
MAI-Voice-2-Flash: A lower-cost variant of the voice generation model, announced at Build with a later launch timeline

Microsoft's technical team described the model family's construction philosophy in a memorable phrase: built on a shared foundation, hill-climbing from the bottom with zero distillation. The zero distillation claim is significant because it means these models were not trained by having them mimic the outputs of OpenAI or Anthropic models, which would be a common shortcut but would also create intellectual property and dependency questions. The MAI models were trained from the ground up on Microsoft's own data infrastructure and evaluation framework, which gives the company genuine ownership of the resulting capabilities rather than a derivative of someone else's work.

"They are designed to work together, and to integrate directly into the products people use every day. But the models themselves are only part of the story."
- Microsoft, at the Build 2026 announcement of the MAI model family

The Excel and Outlook Deployment: What Is Actually Changing

The Bloomberg report that surfaced on July 7, 2026 confirmed something that previous reporting had anticipated but not established as operational: Microsoft's MAI models are now actively handling user prompts in Excel and Outlook, not just in internal testing or limited pilots. The scale, tens of thousands of prompts per week, is meaningful relative to the baseline but represents a small percentage of the total Copilot request volume across Microsoft's 365 user base of hundreds of millions of subscribers.

Microsoft declined to comment on the specifics when contacted by PYMNTS and other outlets, which is consistent with how companies typically handle quiet operational shifts in their inference routing. The company has no business incentive to publicly quantify how much of its AI workload it has moved off external providers, since that number directly reflects its negotiating position with OpenAI and Anthropic. What the Bloomberg report establishes is that the shift is real and operational, not hypothetical or aspirational.

The specific workloads routed to MAI models in Excel and Outlook are not publicly specified, but the pattern visible in other companies that have made similar moves suggests the routing strategy is task-selective rather than random. Simpler, more standardized inference tasks, things like formula suggestions, email summarization, and short text generation, are good candidates for lower-cost in-house models because the quality bar is achievable with models that are cheaper to run than frontier models, and the volume of these tasks is enormous. Complex, long-context, or high-stakes generation tasks, where frontier model quality is more important, continue to be routed to external providers where the capability premium justifies the cost.

GitHub Copilot: The Developer Tool Integration

Beyond the Office 365 productivity suite, Microsoft is integrating MAI-Code-1-Flash into GitHub Copilot, the AI coding assistant that has become one of the most widely used developer tools globally. GitHub Copilot has been powered primarily by OpenAI models since its launch, and the integration of a Microsoft-owned coding model represents the clearest example of how the MAI deployment is about more than just cost management in productivity software.

GitHub Copilot is a commercial product with its own revenue stream and its own cost structure. Every code completion suggestion, every inline explanation, every pull request summary that GitHub Copilot generates comes with an inference cost that Microsoft pays to whoever's model is running the suggestion. At the scale of millions of active Copilot users, the difference between paying external model providers for those inference calls and running equivalent quality responses on owned infrastructure is a substantial financial variable. Microsoft's claim that MAI-Code-1-Flash delivers performance comparable to an earlier-generation Anthropic coding model at lower cost directly addresses this calculation.

The practical experience for GitHub Copilot users is not described as changing in any visible way during the transition. Microsoft is not advertising which model is powering any given suggestion, and the routing between MAI models and external models presumably happens transparently based on task type, context length, and quality requirements rather than on any visible interface change. The user experience stays the same; the model powering it is what changes.

Teams and Transcription: The Next Wave of MAI Integration

Beyond the current Excel, Outlook, and GitHub Copilot deployments, Microsoft has indicated that MAI Transcribe-1.5 will power speech-to-text features in Teams and other products in the coming months. This is a meaningful expansion because real-time transcription at scale is one of the highest-volume and highest-cost AI tasks that a company like Microsoft runs across its enterprise product portfolio.

Microsoft Teams generates an enormous volume of transcription requests across its hundreds of millions of users: meeting transcription, Teams Phone call transcription, voice message transcription, and real-time caption generation for accessibility. The aggregate compute cost of processing that audio at scale is significant, and routing it through external speech-to-text providers or through general-purpose large language models for transcription adds cost overhead that a purpose-built, owned transcription model eliminates. MAI Transcribe-1.5 as an internal model that Microsoft runs on Azure infrastructure means the inference cost stays entirely within Microsoft's own cost structure rather than being paid out to an external provider.

The Financial Logic Behind the Shift

To understand why Microsoft is making this move now rather than two years ago, it helps to understand the specific financial pressure that has been building as Copilot adoption scales. Microsoft has been investing heavily in AI infrastructure while simultaneously seeing the cost of running AI at consumer and enterprise scale grow with every product improvement and feature expansion. The company's AI-related capital expenditure has been running at historic levels, with data center buildout and GPU acquisition consuming an enormous share of the company's investment budget.

At the same time, Microsoft pays licensing fees to OpenAI, Anthropic, and potentially other providers for the model access that powers its products. The exact terms of these agreements are not public, but the scale of Microsoft's Copilot deployment means even a small reduction in per-inference cost translates to hundreds of millions of dollars in annual savings when multiplied across the user base. By owning models that perform comparably to external models on the most common task categories and running them on Azure infrastructure that Microsoft already owns, the company captures both the model licensing margin and the compute margin on the same inference request.

Satya Nadella has framed the broader strategic shift as a move from simply consuming frontier models to participating more fully in the frontier AI ecosystem. That framing positions the MAI models as a competitive capability development rather than purely a cost measure, but the two motivations are not mutually exclusive. A company that builds genuinely capable proprietary models both reduces its external costs and strengthens its competitive position simultaneously.

What This Means for the Microsoft-OpenAI Relationship

The Bloomberg report and the broader MAI expansion raise an obvious question about Microsoft's relationship with OpenAI, a relationship that has been one of the most commercially significant partnerships in the technology industry over the past several years. Microsoft's $13 billion investment in OpenAI and the multi-year exclusive cloud partnership that accompanied it created a deeply intertwined dependency that looked straightforwardly complementary when it was constructed but has become more complicated as both companies have developed competing interests.

Microsoft has consistently maintained that its partnership with OpenAI remains intact and that the shift to MAI models for certain workloads does not represent a break in the relationship. The practical reality is more nuanced: routing a growing percentage of Copilot inference requests to MAI models rather than to GPT models directly reduces the revenue flowing from Microsoft to OpenAI on a per-inference basis. Whether that reduction is material enough to affect the strategic relationship depends on how aggressively Microsoft scales the MAI deployment and which task categories it ultimately routes through its own models.

The commercial structure of the Microsoft-OpenAI relationship includes elements that are not purely per-inference: equity stakes, cloud commitment revenues that come through Microsoft's Azure cloud rather than OpenAI's own infrastructure, and strategic technology sharing arrangements that extend beyond the specific models running in Copilot. A reduction in per-inference licensing volume does not necessarily unwind those broader arrangements. But it does reduce OpenAI's dependency on Microsoft as an exclusive distribution channel and reduces Microsoft's dependency on OpenAI as an exclusive AI capability source, which makes the relationship more transactional and less structurally interdependent than it was at its peak.

The Broader Industry Trend Microsoft Is Joining

Microsoft's move is the latest in a series of similar shifts across the technology industry as companies that were early adopters of external AI model APIs have grown large enough and technically sophisticated enough to build meaningful internal alternatives. The pattern is becoming recognizable: a company starts with external AI providers because building equivalent models is not feasible at the required timeline, scales to a level where the external cost becomes a material financial burden, and begins building internal models for the highest-volume, most standardized task categories while maintaining external provider relationships for frontier capability access.

Apple has been moving AI inference for on-device features toward custom in-house models for several years. Meta built its own LLaMA model family rather than licensing external models, even as it deployed AI at the scale of billions of daily users. Google uses Gemini across its products rather than licensing from external providers. Amazon's Bedrock platform offers multiple model options partly because AWS itself runs its own Titan models alongside the external models it hosts. In each case, the proprietary model is not necessarily the most capable available at the frontier, but it does not need to be for the specific tasks it handles.

What makes Microsoft's version of this trend notable is the combination of its existing OpenAI relationship and the specific timing. Building proprietary models while maintaining a multi-billion-dollar partnership with the company whose models you are partially replacing creates a more delicate navigation challenge than building in-house models from a position of full independence. Microsoft's decision to proceed anyway, and to do so by training models without distillation from OpenAI's outputs, reflects a calculation that the financial and strategic benefits of internal capability outweigh the relationship complications of building it.

What Microsoft 365 and GitHub Copilot Users Actually Experience

For the overwhelming majority of Microsoft 365 and GitHub Copilot users, the expansion of MAI model usage will be invisible in normal operation. Microsoft does not indicate which model is powering any given response, and the routing between models happens at the infrastructure layer without any visible interface element. A user asking Copilot in Excel to explain a formula or suggest a correction experiences the interaction the same way regardless of whether the response is generated by GPT-4o, Claude, or MAI-Thinking-1.

The cases where users might notice a difference, if any, would be at the edges of model capability. If MAI models perform equivalently to the external models they replace on the specific tasks they handle, there is no perceptible difference. If there are capability gaps in edge cases, complex analytical tasks, unusual reasoning requirements, or long-context comprehension, those might surface as subtle quality variations that sophisticated users could detect but that would be difficult to attribute to model differences without controlled testing.

Microsoft's claim that one of its internally developed coding models delivers performance comparable to an earlier-generation Anthropic coding model at lower cost is framed carefully: comparable to an earlier-generation model, not to the current frontier. That framing suggests the routing strategy is designed to match quality requirements to model capability rather than to use the cheapest model for every task regardless of quality outcome. The tasks routed to MAI models are presumably those where earlier-generation equivalent quality is sufficient, while tasks requiring current frontier quality continue to use external models.

What to Watch as the MAI Deployment Scales

The current deployment, tens of thousands of prompts per week across Excel and Outlook, is a beginning rather than a steady state. Several developments in the coming quarters will indicate how aggressively Microsoft is pursuing the in-house model strategy and what its ultimate scope will be.

The Teams and transcription integration represents the next visible deployment milestone. When MAI Transcribe-1.5 launches in Teams, it will be the largest-volume Microsoft AI deployment to run primarily on an in-house model, given the scale of Teams usage and the volume of audio content it processes. That deployment will provide a more material test of whether MAI models can handle mission-critical enterprise workloads at scale without quality regressions that affect user trust.

GitHub Copilot's integration of MAI-Code-1-Flash is the other near-term milestone that will be visible to a large and technically sophisticated user base. Developers who use Copilot intensively have high sensitivity to code quality and are likely to notice if suggestion quality changes in ways attributable to the model switch. Positive reception of the MAI coding model among active Copilot users would provide the strongest possible external validation of Microsoft's internal model quality claims.

The longer-term question is how Microsoft's model capability investment compounds over time. The MAI family introduced at Build 2026 represents the state of the program at a specific point in its development. If Microsoft continues to invest at the rate its Build announcement suggests, the capability of MAI models relative to frontier models from OpenAI and Anthropic will be a moving target. Whether Microsoft eventually develops MAI models competitive at the frontier rather than just competitive with earlier-generation external models is the question that determines whether the in-house strategy becomes a full alternative to external provider dependency or remains a cost optimization layer beneath continued frontier model access from OpenAI and others.