admin The shift marks a notable change for one of the technology industry’s biggest AI spenders. Microsoft has relied heavily on OpenAI’s models to power Copilot features across its software portfolio, while also adding Anthropic technology for selected Microsoft 365 functions. The company is now routing tens of thousands of user prompts each week through its own Microsoft AI, or MAI, model family, a sign that internal systems are being prepared for high-volume enterprise workloads.
The move does not mean Microsoft is abandoning OpenAI. The two companies remain tied through a long-term strategic partnership, cloud infrastructure agreements and deep product integrations. Microsoft has invested billions of dollars in OpenAI since 2019 and remains one of its most important commercial partners. But the operational balance is changing as AI usage spreads from premium demonstrations into everyday office tasks that can generate large and unpredictable compute bills.
Excel and Outlook are central to that calculation. A spreadsheet formula explanation, email summary, meeting follow-up or draft response may not require the most advanced frontier model available. For high-frequency tasks, Microsoft can use smaller or specialised models that cost less to run while reserving more powerful outside models for complex reasoning, coding or creative work. That approach gives the company more control over margins as Copilot moves deeper into paid workplace subscriptions.
Microsoft AI, led by Mustafa Suleyman, unveiled seven in-house models in June covering reasoning, coding, image generation, transcription and speech. The MAI family includes systems aimed at narrower tasks rather than a single all-purpose model designed to outperform every rival. Company executives have presented the effort as part of a broader strategy to build a full AI ecosystem, not merely resell partner technology inside Microsoft products.
Cost is becoming a defining issue for the AI industry. Large language models consume heavy computing resources, with expenses rising as users ask longer questions, upload larger files and use agentic tools that perform multi-step tasks. Enterprise customers often pay through token-based pricing, where input and output length affect charges. That makes profitability difficult to predict when AI features are embedded into widely used software.
Microsoft’s own financial disclosures show the scale of the infrastructure race. The company has continued to lift capital expenditure to support cloud services, AI training and data-centre capacity. Higher spending has supported Azure growth and helped Microsoft sell AI services to large companies, but investors are watching closely for evidence that AI products can produce durable margins rather than simply drive server demand.
The internal model push also gives Microsoft leverage in a market where dependence on one model provider can create pricing and product risks. By using OpenAI, Anthropic, MAI and open-source models across different use cases, Microsoft can match cost, latency, data sensitivity and performance requirements more flexibly. This multi-model approach is becoming common among large enterprises that want to avoid being locked into a single AI supplier.
OpenAI remains central to Microsoft’s premium AI strategy. Its frontier models continue to power advanced Copilot features and remain deeply integrated with Azure OpenAI Service, which gives corporate customers access to OpenAI systems through Microsoft’s cloud. Anthropic also retains a role in parts of the enterprise AI stack, particularly where customers or product teams prefer Claude models for specific language, coding or workflow tasks.
The latest routing changes are therefore more about optimisation than rupture. Microsoft is effectively creating a tiered AI architecture: expensive frontier models for difficult tasks, lower-cost internal models for routine work, and specialised systems for voice, image, coding or transcription. The company’s advantage lies in owning the application layer, the cloud infrastructure and now a growing set of first-party models.
