Microsoft has quietly started swapping third‑party models for homegrown ones inside Excel and Word. Bloomberg reported the change Tuesday. The company once trumpeted OpenAI and Anthropic as the engines behind Office 365. Now it routes a slice of user prompts to MAI — its internal model family — and keeps the rest on external APIs.
The move is not a breakup. Microsoft still pays OpenAI and Anthropic. But the balance has tilted. At Build last month the company unveiled seven new MAI models, among them an agentic coder and a text‑to‑image generator. That launch signals ambition: Microsoft wants to own more of the stack, not just rent it.
Cost is the driver. Inference bills have swollen to levels that make CFOs flinch. A single large‑language‑model call can cost fractions of a cent; multiply that by billions of Office users and the tab rivals a small country’s GDP. Microsoft’s own Azure infrastructure gives it a structural advantage — it owns the GPUs, the data centers, the power contracts. Running MAI on that metal is cheaper than paying a markup to a partner.
This is not a Microsoft‑only story. Amazon has rewritten Alexa’s brain with its own Titan models. Uber replaced a pricey vendor with an in‑house ranking system. Meta funnels more traffic to Llama variants. Accenture told analysts it cut external AI spend by double digits. The pattern is clear: the token‑maxxing frenzy of early 2024 has given way to austerity.
Token‑maxxing — the practice of stuffing prompts with excess context to squeeze better answers — inflated bills without proportional value. Companies now audit every call. They cache responses. They distill big models into smaller ones. They route simple queries to lightweight classifiers. The result: fewer tokens, lower invoices.
Security teams used to block Chinese models on principle. Now procurement asks for quotes. DeepSeek and Qwen offer pricing that undercuts Western rivals by 60‑80 percent. Some enterprises run them in air‑gapped environments. Others accept the risk because the alternative is shutting down AI features entirely. The calculus has shifted from ideological to existential.
Microsoft’s MAI push also rewrites its relationship with OpenAI. The partnership once looked like a marriage; now it resembles a joint venture with a prenup. Microsoft still hosts OpenAI on Azure and integrates GPT‑4o into Copilot. But every MAI deployment reduces dependency. That leverage matters when contract renewals arrive.
Developers notice. The Copilot SDK now exposes model‑selection knobs. A plugin author can pick MAI‑Coder for code completion, GPT‑4o for chat, or a fine‑tuned Llama for domain jargon. Microsoft markets this as flexibility. In practice it’s a pressure valve: if OpenAI raises prices, the traffic shifts overnight.
Investors should watch the margin line. Azure AI revenue grew 31 percent last quarter, but the cost of revenue grew faster. Every percentage point of inference moved to first‑party silicon drops straight to operating income. The MAI rollout is a margin play disguised as a product launch.
Users won’t see a banner announcing the switch. Word still suggests rewrites. Excel still writes formulas. The experience stays the same; the plumbing changes. That invisibility is the point. Microsoft can iterate MAI in production without a public beta, harvesting real‑world feedback at scale.
The broader lesson: AI is no longer a science project. It’s a utility with a meter. Companies that treat it like cloud compute — optimize, reserve, arbitrage — will survive the next funding winter. Those that keep feeding tokens to the highest bidder will watch their runways shrink.
Microsoft chose to build its own power plant rather than stay on the grid. The grid still exists. But the meters now spin slower.