Summary
Microsoft is building its own line of frontier models while keeping its OpenAI partnership intact. A look at what the two companies have actually committed to, and where the strategy points next.
On 2 June 2026, at its Build developer conference, Microsoft did something it had spent years conspicuously not doing: it put its own frontier AI models on stage. In a launch post published the same day, Mustafa Suleyman, chief executive of Microsoft AI, announced a family of seven models built in-house under the company’s MAI brand, spanning reasoning, coding, image generation, transcription and voice.
For a company that had largely resold and built atop OpenAI’s models since 2019, the announcement read as a turning point. Yet on the same platform that now hosts these home-grown models — Microsoft Foundry — Microsoft also continues to distribute the latest models from OpenAI and other developers. The two facts sit side by side, and together they describe a strategy that is less a break than a deliberate hedge.
A partnership renegotiated in two steps
The room for Microsoft to build independently was created not in one move but two.
The first came in October 2025. In parallel announcements, Microsoft and OpenAI confirmed that OpenAI had completed a recapitalisation, converting its for-profit arm into a public benefit corporation with its non-profit foundation retaining control. Microsoft’s stake was formalised at roughly 27 per cent, valued at about 135 billion US dollars. The restructuring also introduced an independent panel of experts to verify any future declaration that artificial general intelligence (AGI) had been reached, and, according to the companies’ statements, gave Microsoft the freedom to pursue AGI on its own or with other partners, reversing an earlier constraint. As part of the arrangement, OpenAI committed to purchase a further 250 billion dollars of Azure services, while Microsoft gave up its right of first refusal as OpenAI’s compute provider.
The second step came on 27 April 2026, when the companies published a further amendment intended, in OpenAI’s words, to provide “long-term clarity”. Under the revised terms, Microsoft remains OpenAI’s primary cloud partner and OpenAI products are still slated to ship first on Azure unless Microsoft cannot or chooses not to support the required capabilities, but OpenAI may now serve its products across any cloud provider. Microsoft’s licence to OpenAI’s intellectual property (IP) runs through 2032 and is now non-exclusive. Microsoft will no longer pay a revenue share to OpenAI, while OpenAI’s revenue-share payments to Microsoft continue through 2030 at the same percentage, subject to an overall cap. Microsoft remains a major shareholder.
Read together, the two amendments loosened the exclusivity that had defined the relationship while preserving its commercial and infrastructural core. Both companies framed the changes as strengthening, not unwinding, the partnership.
The clause that needed defusing
Central to the renegotiation was how the two companies would handle AGI. The original partnership tied Microsoft’s access to OpenAI’s most advanced technology to the question of whether, and when, AGI had been achieved, a determination that carried significant consequences for IP and revenue terms. The restructured framework addresses this directly by handing verification of any AGI declaration to an independent expert panel rather than to one party alone. It is a procedural answer to what had been the single largest source of uncertainty hanging over the relationship.
Building the in-house line
If the contracts created the space, Build 2026 was the demonstration that Microsoft intended to fill it.
According to Microsoft, the flagship of the seven new models is MAI-Thinking-1, a medium-sized reasoning model the company says was trained from the ground up on clean, appropriately licensed data, without distillation from other developers’ models. In Microsoft’s own blind, side-by-side human evaluations, it says the model was preferred to Anthropic’s Sonnet 4.6 and matched leading models on key software-engineering benchmarks. The company is candid that these results are its own; independent verification will come from outside benchmarking over time.
The remaining models target specific layers. MAI-Code-1-Flash, a five-billion-active-parameter coding model, is integrated into GitHub Copilot and Visual Studio Code. MAI-Image-2.5 covers text-to-image generation and editing. MAI-Transcribe-1.5 handles speech recognition across 43 languages, and MAI-Voice-2 generates speech in 15. The models are distributed not only through Foundry and its own products but also via third-party developer platforms including OpenRouter, Fireworks and Baseten, with developers able to tune the model weights themselves.
Underpinning the models is an effort to own more of the stack. Microsoft says it co-designs its models with its own Maia 200 silicon, reporting a 1.4-times efficiency gain, and that its next-generation GB200 cluster is now operational. It has also introduced what it calls Microsoft Frontier Tuning, a method that lets organisations adapt MAI models to their own workflows using reinforcement learning on their own data. As one example, the company says a tuned model for Excel matched OpenAI’s GPT-5.4 while being up to ten times more efficient.
The company has placed this work under a banner it calls “Humanist Superintelligence”, which it defines as advanced systems designed to serve people and organisations rather than replace them, and to remain under human control. The organising metaphor for the lab itself is what Microsoft terms a “hill-climbing machine”, a structure built to improve cycle after cycle as more compute, data and evaluation are applied.
A named partner in healthcare
Alongside the models, Microsoft announced a collaboration with the Mayo Clinic to co-create a frontier AI model for healthcare, combining Mayo’s clinical expertise and de-identified clinical data with Microsoft’s foundational AI capabilities. Microsoft says the model will be deployed first within Mayo Clinic’s own environment and later made available to other organisations through Foundry and, notably, that the resulting model will be owned by Mayo Clinic. The arrangement illustrates a pattern Microsoft appears to favour for sensitive domains: domain-specific models, built with an established institution, with ownership and data control resting with the partner.
The platform paradox
The detail that most complicates any simple “Microsoft versus OpenAI” reading is that Microsoft continues to sell its rivals’ models. Foundry remains a multi-model marketplace, distributing third-party frontier models alongside Microsoft’s own. Independence, in other words, is being pursued at the model layer while the platform stays deliberately open. Microsoft has characterised its portfolio of model providers as a strength rather than a contradiction.
When the partners go public
Microsoft’s position is unusual in one further respect: it holds equity in both of the leading independent model developers. Following OpenAI’s October 2025 recapitalisation, Microsoft’s stake was valued at about 135 billion dollars, 27 per cent of the company. Separately, in November 2025, Microsoft, Nvidia and Anthropic announced an alliance under which Microsoft committed to invest up to 5 billion dollars in Anthropic, at a valuation reported in the region of 350 billion dollars, while Anthropic committed to purchase around 30 billion dollars of Azure capacity and made its Claude models available through Foundry and across Microsoft’s Copilot products.
Both companies have now taken formal steps toward public listings. Anthropic announced on 1 June 2026 that it had confidentially submitted a draft registration statement on Form S-1 to the US Securities and Exchange Commission for a proposed initial public offering (IPO), stating that this gives it the option to go public once the SEC completes its review and that any offering would depend on market conditions; it has not set a share count or price. The filing followed a funding round in late May 2026 that put the company’s post-money valuation at about 965 billion dollars. OpenAI announced on 8 June 2026, roughly a week later, that it had likewise confidentially submitted a draft registration statement to the SEC, while cautioning that it had not decided on timing and that some objectives may be easier to pursue as a private company; it was most recently valued at about 852 billion dollars. Neither company has set a firm date, and reporting has pointed to potential debuts in late 2026 or 2027. All timelines remain subject to SEC review and market conditions.
Microsoft's AI bets, side by side: about 27% of OpenAI (entered near a $135bn valuation) and up to $5bn in Anthropic (at roughly $350bn), plus contracted Azure spend of $250bn from OpenAI and about $30bn from Anthropic — while its own MAI models compete with both.
For Microsoft, two near-simultaneous IPOs by companies in which it holds stakes, and which also compete with its own MAI line, would carry consequences that cut both ways.
The potential upside
The clearest benefit is financial. Public listings would mark Microsoft’s privately held stakes to public prices and, after any lock-up period, make them tradable. A stake entered into OpenAI at roughly a 135-billion-dollar valuation, and an Anthropic position taken in late 2025 at around 350 billion dollars, would already stand to have appreciated considerably against the companies’ more recent valuations — about 852 billion dollars for OpenAI and about 965 billion dollars for Anthropic — and could rise further at a public listing, strengthening Microsoft’s balance sheet and, potentially, helping fund its own capital-intensive build-out.
Public markets would also bring transparency. As listed companies, OpenAI and Anthropic would disclose audited financials, giving Microsoft and the wider market clearer visibility into the economics of frontier AI, including the cost structures behind the Azure commitments both have made. And because Microsoft profits whichever of the two performs better, its dual stake functions as a hedge: it has financial exposure to the success of the model layer regardless of which developer leads, layered on top of the Azure revenue both have contracted to spend and the MAI models Microsoft is building itself.
The potential downside
The same listings would also accelerate the very independence Microsoft’s MAI strategy is designed to offset. Public capital gives OpenAI and Anthropic large, recurring access to funding that does not depend on Microsoft, reducing Microsoft’s leverage over partners that, under the renegotiated terms, are already freer to use rival clouds and to strike their own deals. Better-capitalised competitors at the model layer would compete more directly with the MAI family, which Microsoft itself benchmarks against models such as Anthropic’s.
A public OpenAI and Anthropic would also be answerable to their own shareholders, with their own incentives around pricing, cloud spend and product direction that need not align with Microsoft’s. Holding visible stakes in two direct rivals, which also compete with Microsoft’s first-party models, is a more complex position to manage in public view than in private, and it ties part of the market’s perception of Microsoft’s AI value to the share-price performance of companies it does not control. Dilution is a further consideration: pre-IPO rounds and the offerings themselves could reduce Microsoft’s percentage ownership over time, and lock-up arrangements would constrain when any gains could be realised.
Taken together, the IPOs would crystallise the logic of Microsoft’s hedge while sharpening its central tension: the company would be simultaneously a shareholder in, a cloud supplier to, and a competitor of the two firms whose listings it stands to benefit from.
What Microsoft says it is doing, and the open questions
Microsoft’s own stated rationale is direct. In its June launch post, the company described the effort as being about “long term self-sufficiency for Microsoft and our partners”, tied to an expectation that the compute used to train frontier models will rise roughly a thousand-fold over the next three years.
What the strategy does not yet settle is whether Microsoft’s models can compete at the absolute frontier or whether their pitch is principally one of comparable capability at lower cost — a question Microsoft’s own framing around efficiency and clean data lineage tends to foreground. Other open questions follow from the contracts and the capital markets: what happens commercially if and when the independent panel verifies an AGI declaration; whether the 2032 IP horizon becomes the next negotiating cliff; how OpenAI’s new freedom to use other clouds reshapes the economics that have underpinned Azure; and whether public listings by OpenAI and Anthropic ultimately strengthen Microsoft’s hand as an investor or weaken it as a partner.
On the evidence the two companies have put on the record, the most accurate description is not a separation but a rebalancing, a partnership that now grants each side more room to act alone, while Microsoft builds the capacity to supply its own intelligence rather than rent all of it. The direction is independence. The relationship, for now, endures.
Microsoft’s AI strategy is not an abstract corporate drama. Where Microsoft places its multi-billion-dollar bets, on its own models, on OpenAI, on Anthropic, signals where its products, platforms and priorities go next, and every enterprise that depends on Microsoft has a stake in reading those signals early. If you want an independent take on Microsoft’s investment strategy and what it means for you, talk to us. We don’t sell Microsoft licences or cloud services, so our advice is independent and serves your interests, not Microsoft’s sales motion.