For most of the generative AI era, Microsoft has played a very specific role: the investor and the landlord. It wrote the cheques, supplied the compute, and resold someone else's intelligence. Copilot ran on OpenAI's GPT models. Azure customers reached frontier AI through the OpenAI partnership. Microsoft's AI story was, in a very real sense, OpenAI's AI story with a different logo on the invoice.
That era ended last week. At Build 2026 in San Francisco, Microsoft unveiled seven in-house AI models, headlined by MAI-Thinking-1, its first general reasoning model. Microsoft is no longer just hosting the race. It is running in it. And the way it chose to enter tells you a lot about where the AI market is heading next.
To understand the shift, you have to understand what came before it. Microsoft's bet on OpenAI, starting in 2019 and eventually totalling around fourteen billion dollars, was arguably the best venture-style investment in corporate history. It gave Microsoft exclusive cloud rights to the most famous AI models in the world. If you wanted GPT-4 at enterprise scale, you ran it on Azure. That exclusivity pulled enterprise workloads onto Microsoft's cloud during the most important land grab in a decade, and it gave Copilot a frontier brain that Microsoft did not have to build itself.
But it was also a structural dependency, and an increasingly uncomfortable one. Microsoft's flagship AI products were built on technology it did not own, made by a company with its own consumer ambitions, its own enterprise sales motion, and an increasingly complicated governance story. Every Copilot licence Microsoft sold deepened its reliance on a partner that was steadily becoming a competitor. Regulators in both the US and Europe were circling the relationship, asking whether a fourteen billion dollar "partnership" was really a shadow acquisition. Something had to give.
On 27 April, it did. Microsoft and OpenAI jointly announced a sweeping restructure of their agreement. Microsoft's licence to OpenAI's intellectual property became non-exclusive. OpenAI is now free to sell its models on any cloud, including Amazon and Google. Microsoft stopped paying a revenue share on OpenAI products it resells through Azure, while OpenAI continues to owe Microsoft a capped revenue cut through 2030. And the famous AGI clause, the contractual trigger that would have ended Microsoft's licence if OpenAI declared artificial general intelligence, is gone entirely, replaced by a fixed licence term running to 2032.
Both companies framed it as a maturing relationship, and there is truth in that. But strip away the press-release language and the strategic reality is plain: each side traded away the thing that constrained the other. OpenAI got freedom to run wherever the GPUs are cheapest. Microsoft got freedom to build, ship, and sell its own models without contractual awkwardness or divided loyalty. Six weeks later, it did exactly that.
The Build 2026 launch was not a single experiment. It was a product line. MAI-Thinking-1 is the flagship: a mid-sized reasoning model with 35 billion active parameters in a sparse mixture-of-experts architecture and a 256K token context window, built for multi-step instructions, long-document reasoning, and code generation. It is currently in private preview through Microsoft Foundry, supports function calling, and is compatible with the standard Chat Completions API, which is a deliberate signal that switching to it should be easy.
Around it sit six siblings. MAI-Image-2.5 and its Flash variant handle text-to-image and image editing, and they are already ranking: number 3 on Arena's text-to-image leaderboard and number 2 for image editing, ahead of Google's Nano Banana models. MAI-Code-1-Flash, a 5 billion parameter coding model, began rolling out to GitHub Copilot users the same day. MAI-Transcribe covers 43 languages, and the MAI-Voice models add speech synthesis across more than 15 new languages. Image, voice, transcription, coding, and reasoning: a complete in-house alternative to the OpenAI catalogue Microsoft has been reselling for four years.
Two details in the launch deserve more attention than they got. First, Microsoft says every MAI model was trained from scratch on commercially licensed data, with zero distillation from third-party models, including OpenAI's. That is both a legal posture and a declaration of independence. Second, Mustafa Suleyman's Microsoft AI division branded the effort a "hill-climbing machine", with Suleyman talking on stage about "humanist superintelligence" designed to serve people rather than replace them. Whatever you make of the philosophy, the message underneath it is unambiguous: this is a long-term program, not a one-off release.
Here is the obvious objection. MAI-Thinking-1 has not appeared on public rankings, where Anthropic, Google, Meta, and OpenAI dominate. Microsoft's own performance claims, including that it matches leading models on software engineering benchmarks and wins blind human preference tests against Claude Sonnet 4.6, have not been independently verified. So how does Microsoft plan to compete without the best model on the board?
The answer is that it may not need one, because Microsoft is not playing the leaderboard game. It is playing three different games simultaneously.
Margins. Every GPT token Microsoft serves through Copilot carries OpenAI's economics inside it. Every MAI token does not. A mid-sized 35 billion parameter model is dramatically cheaper to run than a frontier giant, and Microsoft owns the silicon, the data centres, and now the weights. When you sell AI to hundreds of millions of seats, shaving the cost per token is worth more than topping a benchmark. This is the same playbook Microsoft has run for decades: let others pioneer, then win on cost, integration, and scale.
Data. Training exclusively on commercially licensed data is not just caution, it is a sales pitch. Enterprise legal teams are increasingly nervous about IP exposure in AI outputs, with copyright litigation against AI vendors mounting. A model whose training data provenance is clean, and which can be fine-tuned on a customer's own data inside their existing Microsoft tenancy without that data leaving their governance boundary, is precisely what risk-averse enterprises have been asking for.
Distribution. This is the part competitors should worry about most. Microsoft does not need to convince anyone to download a new app. MAI models ship inside GitHub Copilot, VS Code, Microsoft 365, and Foundry, surfaces that hundreds of millions of people already use at work every day. Build 2026 made clear that Copilot and Foundry are becoming orchestration layers where the model is a configurable component underneath. Once the model is a dropdown menu inside a Microsoft product, the default option wins most of the volume, and Microsoft controls the default.
For OpenAI, this is an unideal twist. The company that once wrote the cheques is now building the competition, owns the distribution channel, and no longer pays it a revenue share. OpenAI keeps Microsoft's investment and its Azure capacity, but it has lost its structural moat inside the world's largest software ecosystem.
If your business runs on Microsoft 365 and Azure, and most Australian businesses do, this strategy shift lands directly on your desk. A few things worth thinking about now.
Model choice is becoming a procurement decision. Copilot and Foundry now let you pick the model per workload. Defaulting to whatever ships in the box may mean overpaying for capability you do not use, or quietly accepting a weaker model where quality genuinely matters.
Token cost matters more than benchmarks for most workloads. Summarising documents, drafting emails, classifying tickets, and powering internal copilots do not need the smartest model in the world. They need a model that is reliable, cheap at scale, and compliant with your data governance. A mid-sized model at a fraction of the price can cut AI spend significantly with no visible difference to users.
The licensed-data angle reduces legal risk. If your legal team worries about IP exposure in AI outputs, a topic we have covered before on this blog, training-data provenance is now a real differentiator between vendors, not a footnote.
Take the lock-in lesson seriously. Even Microsoft and OpenAI broke up. Architect your AI integrations so models are swappable, because the leaderboard and the pricing will keep changing, and the businesses that benefit will be the ones that can move.
Microsoft entering the frontier does not mean you should switch to MAI models tomorrow. MAI-Thinking-1 is still in private preview, its benchmark claims are Microsoft's own, and independent verification has not happened yet. But the strategic direction is set, and it will reshape the cost-performance landscape for business AI whether or not MAI-Thinking-1 ever tops a leaderboard. The competition Microsoft just started is not about who has the smartest model. It is about who delivers good-enough intelligence at the lowest cost inside the tools you already use, and that competition benefits buyers who pay attention.
The businesses that win will treat model selection as an ongoing strategic decision rather than a one-time default. If you want help working out which models fit which workloads in your business, and what your Azure bill should actually look like, that is exactly what our AI consulting service is for.
