Microsoft Quietly Swaps OpenAI and Anthropic for In-House AI to Cut Costs

The Shift Begins in Office 365

Microsoft has started routing a meaningful portion of user queries in Excel and Word through its proprietary MAI models, stepping back from its earlier reliance on third-party AI providers. The company had previously highlighted that large sections of Office 365 drew on technology from both OpenAI and Anthropic, but the cost equation is now forcing a rethink.

At DailyTechWire, we've tracked similar moves across the region: the same pressure that drove Seoul's Naver to optimize inference latency and pushed Singapore's Sea to experiment with smaller, task-specific models is now landing in Redmond. The difference is scale. Office 365 serves hundreds of millions of daily users, and even a modest percentage of queries shifted in-house translates to tens of millions of dollars in avoided token charges each quarter.

Microsoft declined to elaborate on the specifics when approached for comment, but the timing aligns with a broader pattern. Last month the company unveiled seven new MAI models at its Build conference, including an agentic coder and a text-to-image generator. Those launches were framed as developer tools, yet their real purpose appears clearer now: they are the foundation for a cost-arbitrage strategy that lets Microsoft retain margin while keeping the AI features users expect.

Why Inference Costs Are Forcing Change

The economics of large-language-model inference have turned punishing. Every time a user asks Copilot to summarize a spreadsheet or rewrite a paragraph, the system consumes tokens. When those tokens are served by OpenAI's GPT-4 or Anthropic's Claude, Microsoft pays a wholesale rate that still adds up fast at enterprise scale. Multiply that by the installed base of Office 365 Enterprise and Government SKUs, and the line item becomes impossible to ignore.

Earlier this year, a brief wave of "tokenmaxxing" swept through parts of Silicon Valley: startups competed to push the highest possible throughput, and model providers raced to demonstrate scale. That phase has ended. The past three months have brought a steady drumbeat of cost-cutting announcements. Amazon, Uber, Meta, and Accenture have each adjusted their AI spending, either by renegotiating contracts, throttling usage, or building internal alternatives.

The sticker shock has grown severe enough that some companies are exploring Chinese foundation models as a cheaper alternative for agentic workflows, despite security concerns that make such arrangements politically fraught. The appeal is straightforward: inference on models from DeepSeek or Baidu can run at one-fifth the cost of comparable Western offerings, and latency from Hong Kong or Singapore data centers is often acceptable for non-critical tasks.

MAI Models as the Arbitrage Play

Microsoft's MAI family is designed for exactly this arbitrage. The models are smaller and more specialized than GPT-4 or Claude, optimized for narrow tasks like formula generation in Excel or style suggestions in Word. They sacrifice some generality in exchange for speed and cost efficiency, and they run on Azure infrastructure that Microsoft already owns. The capital expenditure on GPUs and data-center build-out remains substantial, but the marginal cost per query drops sharply once the infrastructure is deployed.

This is not a wholesale replacement. Microsoft continues to use OpenAI and Anthropic models for more complex prompts, longer contexts, and scenarios where nuance matters. The strategy is hybrid: route the high-volume, low-complexity queries to MAI, and reserve the expensive third-party calls for cases that justify the cost. It is a textbook example of tiering, and it mirrors the approach we have seen in Mumbai and Jakarta, where fintechs use lightweight models for customer-service triage and escalate only the hard cases to frontier models.

The financial logic is compelling. If Microsoft can shift even twenty percent of Office 365 inference to MAI models, the annual savings could exceed several hundred million dollars. That money can then be redeployed into GPU clusters for training the next generation of MAI models, creating a virtuous cycle that reduces dependency on external providers.

The Broader Industry Reckoning

Microsoft is hardly alone. Across Asia and North America, companies that rushed to embed AI features are now confronting the reality that inference at scale is expensive, and the cost curve is not improving as fast as early optimists predicted. The result is a wave of pragmatism. Firms are auditing which features actually drive user engagement, shutting down or throttling those that do not, and investing in smaller, task-specific models that can be fine-tuned and deployed internally.

This shift has implications for OpenAI and Anthropic. Both companies have built business models around selling API access at volume, and Microsoft is one of their largest customers. A meaningful reduction in call volume from Redmond will pressure revenue, even as both firms continue to raise capital at high valuations. The risk is that other hyperscalers follow suit, building their own models and treating third-party APIs as a fallback rather than a default.

For the broader ecosystem, the trend suggests that the first wave of AI integration is maturing. The land-grab phase, in which companies rushed to add "AI-powered" labels to every feature, is giving way to a more measured calculus. The question is no longer whether to use AI, but where it makes economic sense and how to deliver it without eroding margin. That discipline is healthy, but it also means slower growth for model providers and a longer road to profitability for startups that depend on API resale.

What Comes Next

The next chapter will likely see further fragmentation. Large enterprises with the capital and talent to build their own models will do so, carving out niches where they can achieve cost parity or better with third-party offerings. Smaller companies will remain dependent on APIs, but they will negotiate harder and demand better pricing. Model providers will respond with tiered offerings, volume discounts, and partnerships that blur the line between vendor and customer.

Microsoft's move is a signal that the era of unconstrained AI spending is over. The company that bet billions on OpenAI and positioned itself as the AI-first hyperscaler is now quietly hedging that bet, building internal capacity to reduce dependency and protect margin. It is a rational choice, and one that other players in Seoul, Singapore, and Shenzhen are watching closely. The lesson is clear: in the long run, owning the model matters as much as owning the data center.