Once upon a time, OpenAI needed Microsoft.
Today, Microsoft needs to prove it doesn't need OpenAI.
On June 2nd, Build 2026 opened. Microsoft CEO Satya Nadella unveiled seven in-house models, an AI workstation for developers, an enterprise Agent governance platform, and a quantum chip in one go. The information density was off the charts, with every move pointing to the same thing:
Microsoft is parting ways with its closest ally.
But if we zoom out a bit.
You'll find that the situation Microsoft faces today is not unfamiliar.
Thirty years ago, Microsoft defined the PC era. Twenty years ago, the internet rose, Google defined search, Facebook defined social media, Apple defined mobile internet, while Microsoft gradually retreated to the background, becoming the most profitable yet least imaginative infrastructure company of the era.
Later, Satya Nadella spent a decade bringing Microsoft back to center stage.
Today, as AI becomes the new operating system, Microsoft suddenly finds itself standing in that familiar position again.
OpenAI and Anthropic are defining AI, Cursor is defining development, Google is back at the AI table.
Microsoft seems to be selling cloud services again.
I. April 27th, That Crack
The turning point of the story isn't Build itself. It's a deliberately low-key revised agreement from two months prior.
On April 27th, Microsoft and OpenAI simultaneously issued statements announcing their cooperation was entering its next phase. The wording was gentle, the content dramatic:
Microsoft's intellectual property license for OpenAI's models and products changed from exclusive to non-exclusive.
OpenAI can offer products through any cloud service provider, no longer bound to Azure.
Microsoft will no longer pay OpenAI a revenue share.
To put it bluntly, the breakup papers are signed, but they're still living together.
This means the biggest moat Microsoft bought starting in 2019 with $13 billion—that it was the only one globally who could run OpenAI models—was broken overnight.
OpenAI can partner with AWS, with Google Cloud, with anyone.
Microsoft went from being the AI era's exclusive partner to one of its primary cloud service providers.
This is why Build 2026 became Nadella's most crucial keynote. He had to answer a question: Without exclusive rights to OpenAI, what makes Microsoft still the protagonist of the AI era?
II. The Real Problem Behind the Seven Models
The seven models released on Build day—MAI Thinking 1 (Reasoning), MAI Code 1 Flash (Code), MAI Image 2.5 (Image), MAI Voice 2 (Voice), MAI Transcribe 1.5 (Transcription)—cover almost all core capabilities of the AI product chain. This is also the largest simultaneous release of in-house models in Microsoft's history.
But what's truly noteworthy about these models isn't their parameter count or benchmark scores; it's who they're benchmarking against.
Microsoft's AI lead, Mustafa Suleyman, said one thing in an interview: We're focusing more on an Anthropic-style direction: enterprise, developers, and coding.
This directly named the opponent: it's Anthropic, no question.
MAI Code 1 Flash was directly compared to Claude Haiku 4.5, scoring 51.2% on SWE Bench Pro versus Haiku's 35.2%. MAI Thinking 1 is benchmarked against Claude Sonnet 4.6.
Why is the opponent Anthropic and not OpenAI?
The answer is obvious.
According to the Ramp AI Index, in April 2026, Anthropic's enterprise paid adoption rate reached 34.4%, surpassing OpenAI's 32.3% for the first time. In about 70% of initial AI service procurement head-to-heads, enterprises signed with Claude, not ChatGPT. Claude Code holds 54% of the AI programming tools market, while GitHub Copilot has fallen to about 25%.
More awkwardly, Microsoft's own engineers are also using Claude.
According to media reports, Microsoft's internal evaluation found that core developer satisfaction with its own Copilot was lower than with external competitors. The development culture is being infiltrated by external tools.
This is the truth behind the seven models. It's not that Microsoft is also awesome; it's that Microsoft must save itself.
Over the past three years, the entire AI industry believed the model was everything. But at Build 2026, Nadella repeatedly emphasized not one model, but 11,000 models.
Behind this is a classic Microsoft logic.
The future won't have just one super model. Models will become increasingly like databases, servers, and cloud resources—a standard capability.
Microsoft releasing seven models isn't just to prove it can make models. It's also an attempt to make the models themselves less important.
III. Making Money ≠ Being the Protagonist
If you just look at the financial reports, Microsoft's AI business looks great.
In Q3 FY2026, Azure grew 40%, AI business annualized revenue run-rate hit $37 billion, up 123% year-over-year. This is real money.
But behind these numbers lies an awkward truth.
The vast majority of that $37 billion comes from running models *for others*. OpenAI runs on Azure, part of Anthropic's compute is on Azure.
Microsoft is making infrastructure money, not application money.
What about Copilot, which directly faces users?
In the paid AI assistant market, based on Recon Analytics statistics for US paid subscribers, Copilot's market share fell from 18.8% in July 2025 to 11.5% in January 2026, shrinking 7.3 percentage points in half a year, a relative drop of 39%.
In overall workplace office hours, AI tool actual usage time remains consistently at only 1%, showing almost no increase for several consecutive years. There's neither explosive adoption nor a cliff-like collapse; it's stuck on the edge of mainstream workflows.
M365 Copilot reached 15 million paid seats at the beginning of the year, but a Stackmatix survey shows that among employees with product access, the regular active conversion rate is only 35.8%. Many enterprises bulk purchase licenses, but daily employee adoption rates are far below the seat count numbers. Far fewer people are actually using it for daily work than the numbers suggest.
This is the real dilemma Microsoft faces: making money and being the protagonist are two different things.
Azure makes the most money, but ChatGPT and Claude directly own the users. Nadella knows that if Copilot's penetration rate stays in the single digits, Microsoft in the AI era will be like AWS: a massive, profitable, brandless pipe.
Nobody remembers what AWS looks like. But everyone remembers ChatGPT. This is precisely the problem Nadella must solve.
IV. The Subtlety of Jensen Huang's Appearance
One detail at Build is worth noting.
Jensen Huang appeared via video link at the keynote, showing support for the Surface RTX Spark Dev Box. He said the PC is transitioning from personal computer to personal AI.
Two days earlier in Taipei, he had just launched the RTX Spark chip, announcing NVIDIA's formal entry into the PC processor market. Microsoft is his most important partner on the PC front.
But looking closely at this relationship reveals a more subtle conflict.
In the data center market, NVIDIA's biggest customers are Azure, AWS, Google Cloud. Whoever buys the chips is the winner in AI infrastructure. NVIDIA doesn't pick sides; it sells products to everyone.
In the PC market, the RTX Spark chip is allied with Microsoft and MediaTek. The first OEMs are Dell, Lenovo, HP, Asus. On the surface, Microsoft is the beneficiary; every AI PC runs Windows.
But NVIDIA's real goal in entering the PC market isn't to help Microsoft sell Windows; it's to sell local AI compute. RTX Spark's core selling point is 1 petaflop of AI inference capability—running models locally, running Agents locally. If local compute is strong enough, why would users need to go to the cloud? Why would they need Azure?
While helping Microsoft on stage, Jensen Huang is also loosening the foundation of Microsoft's cloud business.
This is the most surreal aspect of today's tech industry: everyone is a partner, and everyone is a potential competitor.
V. Why Build Downplayed Consumers
Looking at the two-day Build 2026 agenda, it almost entirely downplayed the average consumer.
There was no announcement of new Copilot features. No demos of "AI changing your daily life." Even the AI PC concept, which Jensen Huang shouted so loudly about in Taipei two days prior, was deliberately downplayed by Microsoft at Build.
What replaced it were product and capability pitches aimed at developers and enterprises.
Surface RTX Spark Dev Box: 1 petaflop local AI compute, 128GB unified memory, a workstation built for developers.
Agent 365: An enterprise Agent governance platform managing identity, permissions, access control, and compliance, integrating with all major clouds.
MXC (Microsoft Execution Containers): Underlying containers adding safety guardrails to Agents.
OpenClaw on Windows: Enabling personal Agents to securely enter enterprise environments.
The logic behind this is crystal clear: in the consumer market, Microsoft can't win anymore. Tools like ChatGPT, Claude Code, Cursor engage users directly; Microsoft's Copilot can't break in.
But the enterprise market is different. Enterprises need compliance, identity management, auditing, data isolation, multi-Agent permission controls. These are things ChatGPT and Claude can't do; they are inherently personal products.
Nadella's bet is that the AI era needs an enterprise operating system. This operating system isn't the model itself, but the entire platform that lets models run safely and compliantly within enterprises.
Whoever controls this platform controls the entry point for enterprise AI.
This also explains why Build's final headliner wasn't a model launch, but Agent 365—an AI management platform for managing AI.
VI. What Nadella Truly Fears
What exactly is Nadella anxious about?
It's not that Azure isn't profitable, nor that they can't build models, nor that competitors are too strong. He fears that after OpenAI and Anthropic go public, they won't need Microsoft anymore, and he'll be marginalized.
Anthropic secretly filed an S-1 with the SEC on June 1st, valuation $965 billion. OpenAI is also preparing a confidential filing, expected to submit in the second half of the year.
What happens after they go public?
With their own money, they'll use it to buy compute, build data centers. AWS is already deeply tied with Anthropic. SpaceX supplies Anthropic with $1.25 billion in GPU compute monthly, an annual compute bill approaching $15 billion.
When OpenAI and Anthropic no longer need to run models on Azure, how much of that $37 billion AI annualized revenue can Microsoft hold onto?
This is Nadella's real time window.
He must transform Microsoft from being the infrastructure that runs models for others into the enterprise platform all AI must go through, before OpenAI and Anthropic achieve independence.
Models can be swapped. Clouds can be swapped.
But if your identity system, compliance framework, audit logs, and security containers all run on Microsoft's Agent 365, you can't swap out Microsoft.
This is what Build 2026 is really doing: laying a layer of infrastructure only Microsoft can build, beneath all the models.
【Beyond the Layout】's Words:
Another trend revealed by Microsoft Build 2026 is that models are gradually shifting from being the protagonist to becoming infrastructure.
As more and more enterprises use OpenAI, Anthropic, Gemini, and various open-source models simultaneously, what truly matters isn't choosing which model, but who manages these models.
If in 2019, Microsoft leveraged OpenAI to get a first-class ticket to the AI era.
Then in 2026, Microsoft is doing something else. Nadella wants to prove he's not just a passenger, but the driver.
He understands better than anyone that Microsoft once missed mobile internet and once retreated from center stage.
This time, he doesn't want to experience that again.
This article is from the WeChat public account "Beyond the Layout", author: Hua Hua
