Microsoft vs. its own AI suppliers: keep the harness separate

Microsoft closed its fiscal year on July 29 with numbers that make the AI trade look settled: $90.0 billion in quarterly revenue, up 18%, and $35.8 billion in net income, up 31%, per the company's own earnings release. Buried in the same call was the more useful signal for anyone building on top of this stack — Microsoft is now openly selling against the two labs it funds. If you're picking a model layer, the line Satya Nadella used is the one to steal: keep your harness separate from the model.

What actually happened

The financials first, because they set the stakes. Microsoft Cloud revenue hit $59.3 billion, up 27%. Azure and other cloud services grew 43%, with demand still running ahead of capacity — CNBC reports full-year Azure revenue crossed $100 billion for the first time. Full-year revenue was $331.8 billion against $115.9 billion in capital expenditures. Microsoft added 88 data centers over the year.

Then the awkward part. The quarter included a $3.2 billion gain on Microsoft's Anthropic investment, while its OpenAI stake contributed a $480 million net gain in Q4 against $4.96 billion for the full year. Microsoft profits from both labs — and, per TechCrunch's read of the call, spent it pitching its homegrown MAI family against them: MAI Thinking One, MAI Cyber One Flash, and the Maya 200 silicon Microsoft says delivers 40% better performance per watt for those models. Nadella's framing was that Azure carries "the broadest model catalog in the cloud with over 11,000 models," and that the point of a platform is for the customer "to be in control of their own destiny."

He is selling something. He is also right.

Why model portability matters for your business

Here's the trap. You pick a model, then you build the scaffolding around it — the prompts, the tool definitions, the retries, the eval set, the approval gates, the logging. That scaffolding is the harness, and it's where the actual engineering lives. If the harness is welded to one vendor's SDK and one model's quirks, then every repricing, deprecation, or regional restriction becomes a rebuild instead of a config change.

Microsoft's earnings call is the tell: your cloud vendor, your model vendor, and your model vendor's investor can be the same company, competing with itself, and none of that alignment is stable. We've watched Copilot silently swap the model underneath users and export rules re-price whole regions of compute. The specific vendor doesn't matter. The seam does.

Concretely, that means one internal interface for model calls, provider config in environment variables, and an eval suite you can run against any candidate model on your own workload. When a price changes or a model gets restricted, you re-point and re-run the evals. That's an afternoon, not a quarter.

Sources: Microsoft Investor Relations, TechCrunch, CNBC.