Microsoft Isn’t Chasing the Best Model. It’s Buying Down the Cost of Intelligence.
It would be easy to read Microsoft’s MAI model family as another entry in the frontier arms race. That misses the economics. Microsoft already enjoys privileged access to frontier models: its restructured OpenAI agreement extends IP rights to OpenAI’s models and products through 2032, and Microsoft 365 Copilot deliberately supports several providers – Anthropic’s Claude models joined OpenAI’s in September 2025, selectable in Researcher and Copilot Studio. Why build another family? Because AI broke Microsoft’s oldest equation.
Traditional software was gloriously scalable: write Word once, sell millions of licences, and the marginal cost of one more user’s calculation rounded to zero. Generative AI reintroduces a variable cost on every interaction – each prompt consumes compute – and Microsoft’s own filings attribute the decline in Microsoft Cloud’s gross-margin percentage to the cost of scaling AI infrastructure. With Copilot passing roughly fifteen million paid seats by early 2026, per-prompt economics stopped being a rounding error.
Hence the two-front response. On models, MAI-Thinking-1 – announced at Build 2026 – is Microsoft AI’s first reasoning model: a mid-sized mixture-of-experts design with 35 billion active parameters, trained from scratch on licensed data, which Microsoft positions as matching far larger frontier models on software-engineering benchmarks at markedly lower cost; MAI image, voice and transcription models are already inside PowerPoint, OneDrive, GitHub Copilot, and, soon, Teams, and available through Foundry. Beneath the software sits Maia 200, the in-house inference accelerator Microsoft deployed in January 2026 and describes as “engineered to dramatically improve the economics of AI token generation”, claiming 30 per cent better performance per dollar than the newest hardware in its fleet.
Put the pieces together and the model stops being the product; it becomes a component selected dynamically by capability, latency, and cost. That is no longer projection. By mid-2026, Bloomberg reported, Microsoft had begun routing selected Excel and Outlook workloads – tens of thousands of prompts a week – to its own MAI models to reduce its bills from OpenAI and Anthropic, while frontier models continue to serve the requests that genuinely need them. Copilot is quietly becoming a model router. MAI does not have to beat every frontier model on every benchmark; it has to be good enough, often enough, that an expensive third-party model is no longer the default answer to every prompt. The prize is not artificial general intelligence. It is the familiar Microsoft prize: owning enough of the stack that everyone else has to negotiate.
Two sets of implications follow. For enterprise AI buyers: you are no longer purchasing a model but a routing policy – ask vendors which models serve which of your workloads, how routing changes are disclosed, and how quality and data terms move when the mix shifts; benchmark cost per completed task, not list price per token. For strategy leaders in any AI-exposed business, the deeper lesson concerns the marginal cost curve: whoever controls it sets the pace of the market. Build a credible second source – internal or external – for your highest-volume AI workloads before renegotiation, because in the economics of intelligence, leverage belongs to whoever can walk away from the default.
Sources
Personal analysis. Views are mine, not my employer's.