Microsoft is quietly routing tens of thousands of weekly prompts away from OpenAI and Anthropic to its own MAI models — and the strategic logic is more disruptive than the headline suggests.
The MAI Shift — Key Numbers
7
MAI models launched at Build 2026
10K+
Prompts/week now running on MAI in Excel & Outlook
2
Apps where MAI has already displaced third-party models
$0
Suleyman’s target spend on Anthropic inference — eventually
Bloomberg reports that Microsoft has begun replacing OpenAI and Anthropic models with its own in-house MAI models inside Excel and Outlook, with tens of thousands of Copilot prompts per week now running entirely on Microsoft-built AI. The swap is targeted: routine, high-volume tasks — the kind that dominate enterprise productivity apps — are shifting to MAI, while frontier-grade tasks can still route to OpenAI or Anthropic. Microsoft AI CEO Mustafa Suleyman made the commercial logic explicit: “We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost.”
The MAI lineup — introduced at Build 2026 — spans seven models covering reasoning, coding, image generation, speech, and transcription, led by MAI-Thinking-1 for complex reasoning tasks. This is not a skunkworks experiment. Microsoft is shipping these models into the two most widely deployed enterprise software products on earth.
The nuance that matters: Microsoft is not terminating its partnerships with OpenAI or Anthropic. Copilot and Azure are being architected as multi-model platforms that route each prompt to the best-fit — and often cheapest good-enough — model available. Frontier reasoning tasks can still land on GPT-4o or Claude. What’s shifting is the volume underneath: the enormous, repetitive, commodity inference load that actually drives third-party API bills.
The key insight: Microsoft is not abandoning the frontier labs — it is demoting them from default infrastructure to premium option. That distinction is everything for OpenAI’s and Anthropic’s long-term revenue models.
Microsoft’s Stack Capture — Timeline
2019–2023
Microsoft commits $13B+ to OpenAI; becomes primary distribution and cloud partner. Full dependency on OpenAI models for Copilot.
2024
Microsoft debuts Maia 100 — its own AI training chip — signaling the intent to escape Nvidia’s “chip tax” on inference costs.
Microsoft launches seven MAI models — reasoning, coding, image, speech, transcription. MAI-Thinking-1 targets complex inference. The model stack is now internally sourced.
July 2026
Bloomberg confirms MAI is live inside Excel and Outlook, routing tens of thousands of prompts per week. Suleyman publicly names Anthropic cost elimination as a strategic goal.
The Structural Read
This is the Model Tax story — and it rhymes exactly with the Chip Tax story Microsoft already solved.
When Microsoft built Maia, the logic was simple: Nvidia captured enormous margin on every GPU cycle powering Microsoft’s cloud. The only durable escape was vertical integration — own the silicon, own the cost structure. Maia was that escape hatch on the hardware layer. MAI is the identical move one layer up. OpenAI and Anthropic have captured margin on every frontier inference call routed through Azure and Copilot. As long as Microsoft had no credible internal model, it had no negotiating leverage — it was a price taker.
Now it has one. And critically: it doesn’t need MAI to beat GPT-4o or Claude at frontier reasoning. It only needs MAI to be good enough for the enormous commodity-inference base — the repetitive summarization, drafting, and formula tasks that dominate Excel and Outlook usage. Those prompts don’t need frontier intelligence. They need speed and low marginal cost. MAI delivers both.
The Model Tax Doctrine
The buyer becomes the competitor the moment it has a good-enough alternative.
Frontier labs priced their models on the assumption that no single customer could replicate them. That assumption held — until Microsoft, Google, and Amazon all reached sufficient engineering scale to build internally. The “model tax” follows the same structural logic as the “chip tax”: pricing power erodes exactly when the hyperscaler’s internal option crosses the good-enough threshold for volume workloads. It never needs to win on benchmarks. It just needs to win on margin.
The deeper dynamic is one the AI economy has been building toward since 2023: the Subsidized AGI Economy. OpenAI and Anthropic have grown on the back of hyperscaler distribution — Microsoft’s Office install base, Azure’s enterprise reach, Amazon’s Bedrock, Google’s cloud. That distribution came with a hidden dependency. The distributors now have the usage data, the infrastructure, and — as of Build 2026 — the models to route volume wherever they choose.
Mustafa Suleyman — Microsoft AI CEO
“We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost.”
Three Implications
FOR OPENAI AND ANTHROPIC — REVENUE EROSION FROM BELOW
The frontier labs won’t lose the headline partnership agreements. They will lose the volume. High-volume, low-complexity inference — the workload that actually scales API revenue — is precisely what MAI targets. Losing that volume to Microsoft’s internal routing doesn’t show up as a broken partnership; it shows up as a revenue curve that grows slower than the underlying model adoption curve. That’s a worse problem to diagnose and defend against.
FOR ENTERPRISE BUYERS — THE MULTI-MODEL PLATFORM IS THE PRODUCT
Microsoft’s multi-model routing architecture — where Azure and Copilot dynamically assign each task to the best available model — is itself a competitive moat. Enterprise customers don’t need to pick a model; they buy the platform and the platform optimizes cost and quality automatically. This locks enterprises into the Microsoft stack more deeply than any single model partnership could. The abstraction layer is the advantage.
FOR THE BROADER AI STACK — COMMODITIZATION ACCELERATES
When Microsoft publicly names cost elimination of a frontier-lab partner as a strategic goal, it sends a signal to every other hyperscaler and large enterprise running significant AI inference: build or route internally where good-enough models exist. The era of frontier-model pricing power over commodity-level tasks is ending. What survives is pricing power over genuinely hard problems — long-horizon reasoning, novel research, complex agentic workflows — where MAI-class models don’t yet compete. That frontier is narrower than the labs would like.
Business Engineer Framework
The Map of AI — Where Microsoft Just Moved
The Map of AI tracks 200+ companies across 9 layers of the AI stack — from silicon to applications. Microsoft’s MAI move is a textbook vertical integration play: own chips (Maia), own models (MAI), own distribution (Office/Copilot/Azure). Understanding which layer each player controls — and which layers are being commoditized — is the only way to read who wins the next five years of the AI economy. This framework maps it precisely.
Microsoft is not defecting from the frontier AI ecosystem — it is restructuring its position within it. By owning the chip layer with Maia and now the model layer with MAI, it has built the infrastructure to route commodity inference internally while preserving frontier partnerships for genuinely hard tasks. The model tax, like the chip tax before it, gets eliminated through vertical integration — not negotiation. For OpenAI and Anthropic, the risk is not a broken contract; it is a distribution partner that is systematically shrinking the volume of work they need to do while retaining all the credit for AI capability. That is a much harder problem than losing a deal.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
