Microsoft’s “Super App” Takes on OpenAI and Anthropic Later This Year
Microsoft has posted record numbers for the fourth quarter of its 2026 fiscal year (ended June 30): $90 billion in revenue and $35.8 billion in net income for the quarter, and $331.8 billion in revenue with $133.7 billion in net income for the full year. Growth is being driven by the cloud and AI businesses – while the Xbox division, for instance, declined by ten percent.
The more telling detail, though, sits in the supplementary materials to the earnings report: how Microsoft’s stakes in the two largest – and mutually competing – AI labs have developed.
A $3.2 Billion Paper Gain on the Anthropic Stake
For the quarter, Microsoft booked its Anthropic investment as a $3.2 billion gain, lifting diluted earnings per share by 33 cents (reported diluted EPS for the quarter: $4.81). Microsoft invested $5 billion in Anthropic in November 2025 – part of a circular arrangement under which the AI lab committed to purchasing $30 billion worth of Azure services. Unlike its OpenAI stake, Microsoft does not routinely revalue this investment each quarter.
At OpenAI, where Microsoft holds roughly 27 percent, the quarter went the other way: the company marked the stake down by about $600 million, cutting diluted EPS by roughly seven cents. On a full-year basis the picture looks considerably friendlier – a $5 billion gain and a contribution of $0.67 per share (full-year EPS: $17.95). Microsoft additionally receives revenue-share payments under the OpenAI deal, the size of which it does not disclose.
Measured against Microsoft’s overall results, the OpenAI write-down is a rounding error. What stands out nonetheless is that the paper gain on Anthropic in a single quarter came close to matching the full fiscal year’s gain on OpenAI – and that Microsoft chose to disclose it explicitly.
Nadella’s Message: Keep the Harness Separate From the Model
On the analyst call, Nadella sharpened an argument he has been making for months: enterprises should use multiple models and should not hand the agentic layer – the “harness” – to the model providers. Asked by UBS analyst Karl Keirstead about the open- versus closed-source debate, he said the goal is for firms to stay in control of their own destiny. Architecturally, that means keeping the harness separate from the model so that any model remains swappable at any time.
The argument targets two enterprise IT anxieties: data leakage and vendor lock-in. It also targets a business model – because OpenAI and Anthropic are currently expanding into applications and agentic infrastructure, precisely the layer through which customer relationships are controlled.
As evidence, Nadella pointed to the previous week’s incident at Hugging Face. An unreleased OpenAI model had broken out of its sandbox and – in pursuit of beating a benchmark – mounted an attack on Hugging Face. During the response effort, an unnamed proprietary frontier model initially refused to help; Hugging Face then turned to the Chinese open-source model Z.ai GLM 5.2 to analyse logs. You cannot make yourself dependent on a single model’s refusal, is Nadella’s reading. The incident has rattled the industry broadly – OpenAI CEO Sam Altman is now also talking about potentially slowing the pace of development.
Own Models, Own Chips, One “Super App”
In parallel, Microsoft is expanding its own portfolio. Nadella pointed to a model catalogue of more than 11,000 models – including those from OpenAI, Anthropic, Mistral and xAI, but also the in-house MAI family. More than a dozen new models were announced across image, voice, transcription, coding and security, including the company’s first reasoning model, MAI Thinking One. The models are being co-designed with Microsoft’s own silicon; on the in-house Maya 200 chip the company cites 40 percent better performance per watt.
Aimed directly at Anthropic’s top-tier Mythos model line is MAI Cyber One Flash, unveiled this week. Combined with Microsoft’s multi-agent security harness, it delivers better performance than the much larger Mythos model at half the cost, according to Nadella. Independent comparative testing supporting those claims is not yet available.
At the application layer, Microsoft is bundling its offering: a Copilot “super app” is due this year, merging chat, coding, Cowork and the agentic Autopilots into a single product – spanning both consumer and commercial scenarios. OpenAI is pursuing a similar bundling pattern with its ChatGPT Work app, which combines ChatGPT with its coding tool Codex.
Partner and Competitor at Once
Nadella continues to recommend that enterprises include frontier models from OpenAI and Anthropic in their mix – Microsoft earns from their cloud consumption just as it does from the equity stakes. But the strategic message to enterprise customers is simultaneously: keep your dependence on any single provider low. For Microsoft, that is a comfortable position as long as the two roles do not cannibalise one another. Whether it holds will likely depend on how quickly OpenAI and Anthropic place their own application layer directly with customers – and how well the MAI models actually perform in real-world comparison.
