# Microsoft can fire its model supplier. Can you? At Build 2026 Microsoft shipped its own coding and reasoning models — trained from scratch, with what its AI chief called 'zero distillation' from OpenAI — straight into GitHub Copilot. The richest software company on earth just spent billions to stop depending on one supplier. That's the whole lesson for the rest of us, and it costs you nothing: never let the model be the part of your system you can't swap out. - **Author:** Fedor Chishchin - **Published:** 2026-06-06 - **Tags:** architecture, ai-native, business - **Source:** https://fedorthinks.com/blog/fire-your-model-supplier --- On June 2, on the opening morning of Build 2026, Microsoft did something that would have sounded unthinkable a year ago: it shipped its own models to compete with the company it has poured over thirteen billion dollars into. Two of them — MAI-Thinking-1, a reasoning model, and [MAI-Code-1-Flash](https://microsoft.ai/news/introducingmai-code-1-flash/), a small coding model — and the coding one started [rolling out the same day to every paying GitHub Copilot user](https://www.cnbc.com/2026/06/02/microsoft-unveils-new-ai-models-lessen-reliance-on-openai-lower-costs.html). The detail that matters isn't the benchmark scores. It's how they were built. Microsoft's AI chief Mustafa Suleyman said the models were trained from scratch with ["zero distillation"](https://www.techtimes.com/articles/317631/20260602/microsoft-build-2026-mai-thinking-1-first-house-reasoning-model-trained-without-openai-data.htm) — no OpenAI weights, no OpenAI data, none of the usual shortcut of copying a bigger model's homework. Microsoft built its own from clean, licensed data because it wanted something it didn't have before: the ability to walk away. ## Why the biggest player on the board wanted out For years, Microsoft's entire AI story ran through OpenAI. Copilot, Azure, the lot. That's a lot of leverage to hand one partner. So this year Microsoft renegotiated, stood up its own research team, and started building. Suleyman was blunt about how recently the leash came off: ["we were only sort of set free from our contract with OpenAI about six months ago"](https://venturebeat.com/technology/microsoft-ai-chief-says-company-was-set-free-from-openai-to-pursue-superintelligence). Satya Nadella framed it as a principle, on stage, in front of everyone: > "We believe the time has come for every company to just move from consuming a > frontier model to fully participating at the frontier." Read that again, because it isn't really about Microsoft. _Consuming_ a model means you rent it, you depend on it, and your costs and your roadmap move when someone else decides they should. The most resource-rich software company alive looked at that arrangement and spent a fortune to get out of it. ## You are renting too Here's the uncomfortable part. If you're building a product on an AI model right now, you are doing exactly the thing Microsoft just paid billions to stop doing. You're renting your most important component from a supplier who sets the price, the rate limits, the deprecation schedule, and the terms — and who can change any of them on a Tuesday. We just watched [GitHub flip its own pricing to a meter](/blog/the-token-bill-came-due) and send power users' bills up tenfold. Renters don't get a vote. You can't out-spend Microsoft and train your own frontier model. You don't need to. The lesson isn't "build a model." It's the thing the model independence _buys_ you: **the freedom to swap the supplier without rewriting your product.** And that freedom is almost free. It's an architecture decision, not a budget. ## The seam that lets you fire anyone The mistake I see constantly is a model's name and quirks wired straight through the codebase — this exact API, this exact prompt format, this exact response shape, referenced from forty places. That product can't switch models. Every new release is a migration project, every price hike is a crisis, and the supplier knows it. The fix is boring and it works: put the model behind one clean seam — a single adapter your code talks to, with the vendor's specifics hidden on the far side of it. My own products do this. The language model is a config value; the [engine that computes the real answers](/blog/cheap-model-most-of-the-work) is mine and never moves. When a better or cheaper model ships, I change one line and the product gets better. When a supplier raises prices or kills an endpoint, it's a shrug, not a fire drill. That's the same capability Microsoft just bought, scaled down to your size. They got it by training models on their own clusters. You get it by writing about thirty lines of indirection. The strategic position is identical: _nobody owns the part of your system you most depend on except you._ ## The real question Strip away the billion-dollar numbers and Build 2026 was a company making sure it could fire its supplier. That's the move worth copying — not the training run. So ask the only question that matters about your own stack: if your model provider doubled its price tomorrow, or retired the exact model you depend on, how long would it take you to move? If the answer is "an afternoon," you've already won the thing Microsoft spent billions on. If the answer is "we can't, really" — then you don't have a moat, you have a landlord. And the whole point of this week's news is that even the landlord's biggest tenant decided that wasn't good enough.