Microsoft is panicking.

Fortunately, I stayed up late last night to go through the content of the Build Conference. While the impression is still fresh, let me share my thoughts with you. First, let me mention a change you might not have noticed.

01

Build 2026 only lasted for two days, from June 2nd to 3rd, at the Fort Mason Center in San Francisco. On the first day, Nadella gave the keynote speech, which lasted for two and a half hours. On the second day, there were technical breakout sessions and hands-on labs.

Two days sounds normal, right? It was four days last year, but this year it was cut in half.

Not only was the number of days reduced, but the location was also changed. It was moved from Microsoft's headquarters in Seattle to San Francisco. The official explanation is that it is "closer to the core of the global AI industry."

The number of participants dropped from more than 5,000 to 2,500. And it's not just a matter of paying to get in; you have to apply and go through a review process. The ticket price is $1,099.

The slogan Microsoft came up with is also quite interesting:

Two days. Hands-on sessions. No fluff; Two days, hands-on sessions, no nonsense.

It sounds decent. When you put all these changes together, the signal it sends is that this conference is shrinking. When a company voluntarily scales back its most important annual developer event, it usually only means one thing: It's reorganizing itself.

In my opinion, Build 2026 took place at one of the weakest moments for Microsoft in recent years.

Nadella stood on stage for the entire keynote speech. He talked about the wonderful future where AI agents will do your work for you, from synchronous assistants to asynchronous colleagues, and let agents perform long-term tasks in key areas.

The keywords on stage were "autonomy," "making decisions for you," and "the future is here."

What was the real context off stage?

Microsoft's stock price has dropped by more than a quarter from last year's record high. After the second-quarter earnings report was released in January this year, the stock price plummeted by more than 10% in a single day. Overnight, $440 billion evaporated from its market value.

The old dogs on Wall Street are asking a very straightforward question: You're spending $190 billion a year on building AI infrastructure. Where's the return?

Then there's Copilot, which is Microsoft's most heavily invested and most hyped AI product in the past two years. Nadella keeps saying everywhere that it will change the way every knowledge worker works.

What's the actual situation?

Out of the 400 million Microsoft 365 enterprise licenses globally, only 20 million have Copilot enabled. That's 5 out of every 100 users.

What's even more ironic is the timeline. On June 1st, the day before the Build Conference opened, Copilot experienced a large-scale outage. More than 2,600 people reported problems on Downdetector. Those enterprise users who had reconfigured their workflows around this AI assistant were locked out.

Then, the next day, Nadella stood on stage and told the world that in the future, AI agents will handle all your work.

If you were in the audience, what would you think?

The market's reaction was also very straightforward. On the day of the Build Conference (U.S. time), Microsoft's stock price didn't rise; instead, it fell by 3.6%. It did rise by more than 5% the week before the conference. It's the classic case of buying on the expectation and selling on the fact.

Wall Street is telling you with real money that they've already priced in your story. Now they want to see the real thing.

So, Microsoft Build 2026 is, on the surface, a technological feast and a two-and-a-half-hour future show by Nadella. If you look at the background, it's more like a person being questioned from all sides stepping onto the stage to try to prove themselves: Prop me up, I'm still good.

02

Nadella spent a long time on stage talking about one word: models. Microsoft's own models.

At this Build Conference, Microsoft launched seven self-developed AI models all at once, all under the MAI brand:

MAI-Thinking-1 is responsible for reasoning, MAI-Code-1 for programming, MAI-Image-2.5 for image generation, MAI-Voice-2 for voice synthesis, and MAI-Transcribe-1.5 for voice-to-text conversion.

It covers five areas: reasoning, programming, image, voice, and transcription.

This scale is quite unusual. In the past few years, when Microsoft talked about AI, it was always Copilot in the front and OpenAI in the back. When you used Copilot, it was running on GPT under the hood.

Now Nadella suddenly puts "Microsoft's own models" in the spotlight and specifically adds:

MAI-Thinking-1 was trained from scratch without using the output of any other AI system. There was no distillation.

Think about it. What this means is that this model has nothing to do with OpenAI. So why is Nadella emphasizing this at this time? We need to look back at the calendar.

On April 27th, Microsoft and OpenAI revised their cooperation agreement. There are three core points:

First, Microsoft's authorization for OpenAI's models changed from exclusive to non-exclusive. What does that mean? OpenAI's products can now run on any cloud; they don't have to use Azure.

Second, Microsoft will no longer share revenue with OpenAI. Third, these terms are in effect until 2032, but the exclusive partnership has been terminated.

In simple terms, this "marriage" has changed from "you can only be with me" to "you can date others, but the house is still mine." Then things got even more interesting. Not long after the agreement was revised, OpenAI turned around and signed a huge deal worth $50 billion with Amazon AWS.

Who is AWS?

It's Microsoft Azure's biggest competitor in the cloud computing market. You know, OpenAI owes a lot of its success to Microsoft. Microsoft is the real benefactor. From 2019 to now, Microsoft has invested more than $13 billion in total.

Microsoft fed Azure's computing power to OpenAI, allowing it to train GPT-3 and GPT-4. Now OpenAI is taking the models trained with Microsoft's money and giving the business to Microsoft's biggest competitor.

You can probably guess what Nadella is thinking.

There's an interesting report from Agence France-Presse. Nadella has been saying for years that he refuses to let Microsoft become another IBM.

Who is IBM? It's the computing giant that gave the PC operating system contract to a small company called Microsoft, helped it grow, and was eventually left behind.

History can be so neat. Back then, IBM gave Microsoft a hand, and Microsoft overtook IBM. Now Microsoft has helped OpenAI, and OpenAI is breaking free from Microsoft.

Nadella obviously doesn't want to be this generation's IBM.

So, launching seven self-developed models at the Build Conference is Microsoft telling the market that even if OpenAI leaves tomorrow, it can still do it on its own.

Sophie Lebrecht from Microsoft's AI team said something very straightforward during a media tour before the Build Conference:

It's important that we are self-sustaining and are not taking huge dependencies; It's important to be self-sufficient and not overly reliant on any single external force.

What's even more notable is that I found out:

Microsoft's self-developed programming model, Polaris, will replace GPT-4 Turbo as the default model for GitHub Copilot in August this year. That means the underlying model of Microsoft's most core AI developer tool will be replaced with its own.

This is preparing for a future without OpenAI. And the next question Nadella has to face is: Are you ready? Can you really win in such a competitive market?

03

I don't have an answer. I always feel that whenever someone shouts "reinvent," the ending is never very good.

At this Build Conference, besides Nadella, another person stole the show. Jensen Huang.

The day before the Build Conference, Jensen Huang released the RTX Spark chip at Computex in Taipei. Then he appeared on the Build Conference stage. The two conferences were almost simultaneous, and the same product was launched. It was a coordinated effort.

Jensen Huang said something very significant at Computex: This reinvention of the computer is as significant as the reinvention of the mobile phone into the smartphone.

"Reinvention." Reinvention.

This word carries a lot of weight in Silicon Valley. It means: We're creating a new category. This is the framework Steve Jobs used when he launched the iPhone in 2007.

But do you know? The people who shouted "reinvent" in the context of Microsoft in the past didn't have a good ending.

In 2012, Microsoft launched the Surface RT. It was Microsoft's first attempt at making hardware, with the goal of "redefining the tablet." They wanted to beat the iPad with a magnetic keyboard.

Microsoft executives were very confident at the launch event, thinking it was a moment to rewrite PC history.

Sinofsky, the then-head of Windows, sent an internal email to CEO Ballmer a few weeks after the Surface RT was launched. The original text was like this:

We're really in a very tough spot and I feel like we're on the verge of an unrecoverable situation; We're in a very difficult situation, and I feel like we're on the brink of an irreversible situation.

Ha ha, he was right.

The Surface RT was priced starting at $500, and you had to pay an additional $100 to $130 for the keyboard, which was its biggest selling point. For $500, you could buy an iPad, which had hundreds of thousands of apps, while the Surface RT had almost none.

Microsoft overestimated consumers' tolerance for the Windows brand premium and its own judgment of the supply chain.

The final result:

Six million tablets were left unsold in the warehouse. Microsoft took a $900 million inventory write-down. The day after the announcement, the stock price plummeted by more than 10%, and $30 billion in market value evaporated.

I checked the history. The Huffington Post later reported that there were millions of beautiful but unsold Surface RT tablets piled up in an unknown warehouse. They were supposed to "revive Microsoft."

This wasn't the end. In September 2013, not long after the failure of the Surface RT, Ballmer decided to acquire Nokia's mobile phone business for $7.2 billion.

The logic was clear: The Windows Phone operating system needed its own hardware to compete with Android and iOS. Just like Apple, it wanted to integrate hardware and software.

Nadella opposed this deal at the time, but Ballmer pushed it through.

Later, Ballmer left, and Nadella took over as CEO. In 2015, Nadella took a decisive action: He almost completely wrote off the Nokia acquisition, wiping out $7.6 billion directly. He laid off 7,800 employees, most of whom were from the mobile phone department.

This was the biggest write-down in Microsoft's history, even bigger than the $6.2 billion write-down of aQuantive before.

Later, analyst Jack Gold said something that was repeatedly quoted by major media: This was a mistake from the beginning, a huge mistake. Microsoft should never have entered this low-margin, cutthroat mobile phone market.

Two "reinventions," two failures.

The total cost: $900 million plus $7.6 billion, nearly $8.5 billion down the drain. Nearly 8,000 people lost their jobs. Two generations of hardware product lines were axed.

So, when Jensen Huang stands on stage now and says "reinvent the PC," we have every reason to ask: Is this the iPhone moment, or another Surface RT moment?

04

This question is really worth thinking about seriously because the answer isn't one-sided.

Microsoft doesn't have a history of all failures. Nadella himself won a major "reinvention."

When he took over as CEO in 2014, Microsoft was at a low point. During Ballmer's tenure, Microsoft missed out on the smartphone market, failed with tablets, and even lost market share in the PC operating system. The stock price dropped by 40% during Ballmer's term.

Nadella did one thing when he took over:

He transformed Microsoft from a company that sold Windows licenses to a company that sold cloud services. Azure.

He bet right. The demand for enterprises to move to the cloud was real. Amazon AWS had been in the market for several years, the market had been educated, and customers were already moving to the cloud on their own.

All Microsoft had to do was use its strongest weapons, its enterprise relationships and full suite of office software, to grab market share in an existing trend.

By the third quarter of this year, Azure's revenue increased by 31% year-on-year. More than 80% of the Fortune 500 companies are using Azure AI services.

Nadella took ten years to transform an "outdated PC company" into a $3 trillion cloud giant. So, the real question at Build 2026 is "Is the demand really there?"

The Surface RT failed because it was creating a product for a non-existent market. Azure succeeded because the demand was already there. Which category does the AI agent economy fall into?

If you look at the data, the evidence is contradictory.

Nadella says that agents are the next platform-level opportunity, and the whole conference revolved around this. But the Copilot data I mentioned earlier shows that even for the relatively simple form of an "AI assistant," enterprises aren't buying in on a large scale. Now, are we going to jump to "AI colleagues working autonomously"? There's still an unclear catalyst between seeing the potential and actually spending money.

However, the situation this time is really different. There's an interesting detail.

In 2012, the Surface RT used NVIDIA's Tegra 3 chip. That's right, NVIDIA and Microsoft worked together on Windows on Arm back then and failed. Huang left this ecosystem for several years.

Now in 2026, the two companies are back on the same stage, promoting the RTX Spark. The alliance structure is completely different. During the Surface RT era, Samsung, Lenovo, ASUS, and Dell withdrew one after another, and in the end, only Microsoft and Nokia were left selling the product.

What about the RTX Spark?

The first batch of OEMs includes ASUS, Dell, HP, Lenovo, MSI, and Microsoft Surface. Acer and Gigabyte are following closely. More than 30 laptops and more than a dozen desktop devices are in development.

Huang also announced a three-generation roadmap:

After Spark comes Vera Rubin, and then Rosa Feynman. It's like telling the entire industry chain: I'm not going anywhere this time.

But "a luxurious lineup" doesn't mean "a sure win."

NVIDIA has two mountains to climb. One is Apple. The M series has dominated the high-end Arm laptop market since 2021, and NVIDIA is coming too late. The other is memory. There's a severe global shortage of AI memory, and the hardware supply is uncertain. How can it support a new platform?

Moreover, in front of all these technical problems, there's an even more fundamental psychological problem:

Do users really want an "AI colleague" to help them with work, or an "AI boss" to make all the decisions for them? Just because technology can do it doesn't mean people are willing to accept it.

Nadella, the old guy, doesn't seem to be completely certain either. At the end of the Build Conference, instead of coming up with a catchphrase like at a typical tech conference, he threw out a topic:

The most terrifying possibility of AI is that technology centralizes power and reduces humanity. He said that Microsoft is trying to build an alternative