What Was Announced at Microsoft Build 2025?
Microsoft Build 2025 was one of the more satisfying developer conferences in recent memory — and not just for the announcements. The company showed up with its own foundation models, a new AI agent deeply integrated into Microsoft 365, a vision for a completely new device category, and a bold embrace of open-source AI. Whether you're a developer, enterprise buyer, or just someone trying to understand where AI is heading, Microsoft Build 2025 had something worth paying attention to. Here's everything that was announced, ranked by what actually matters.
The headline themes were efficiency, agents, and platforms. Microsoft isn't trying to win the frontier model race on raw performance — it's trying to win on cost-per-token ROI, enterprise trust, and ecosystem lock-in. That's a different game, and Build 2025 was the clearest statement yet of that strategy.
What Is Microsoft Project Solaris and Why Does It Matter?
Project Solaris is Microsoft's most ambitious and polarizing announcement from Build 2025. It's a new Android-based operating system designed to run agents instead of apps — and it comes with two new hardware form factors to go with it. The first is a stationary desk device built on MediaTek silicon, think concept car energy, with Windows Hello for business letting you walk up and sign in securely. The second is arguably more interesting: a lightweight portable device that looks less like a phone and more like a smart badge or key card you'd wear. It has a screen, supports fingerprint unlock, and is designed specifically for firing off agent tasks on the go.
08:15
Satya Nadella introduces Project Solaris, showing the stationary desk device and portable badge-form-factor hardware
Watch at 08:15 →
Analyst Ben Thompson called it vaporware — but compelling vaporware. Microsoft has signed up Qualcomm and MediaTek as chip partners and showed real physical devices on stage, so it's not purely theoretical. What Thompson found most interesting wasn't the hardware itself but the model it represents: the cloud as the hub, with multiple devices as spokes, instead of the phone being the center of everything.
His argument is that agents work best in the cloud. They need to reach across apps, data, and devices. The phone, locked down by Apple or Google, is actually a bad hub for that kind of work. A thin client — even a badge-sized one — that securely connects an employee to their enterprise agents running on Azure and Microsoft 365 makes a lot of sense in an enterprise context. Companies could mandate employees carry them. The badge becomes a secure on-ramp to cloud-native AI workflows.
The comparison to the Rabbit R1 came up immediately. That device got mocked, but the core idea — offload everything to the cloud, have a minimal client for interfacing — looks a lot smarter now that agents are actually capable of doing real work in the background. The key insight: the device's usefulness isn't in what it displays. It's in what the agent is doing in the cloud while you've already moved on.
12:40
Demo of the Project Solaris portable device — fingerprint unlock and agent task queue on a badge-sized screen
Watch at 12:40 →
What Is Microsoft Scout and How Does It Use OpenClaw?
Microsoft Scout is the company's first proactive AI agent for Copilot, and it's a big deal for a reason that almost got buried in the keynote: Scout will be powered by OpenClaw, the open-source AI agent framework. When Scout rolls out more widely this summer, Microsoft will also contribute its security guardrails back to the OpenClaw open-source ecosystem.
Scout operates across cloud, desktop, and web — connecting to Teams, Outlook, OneDrive, SharePoint, and all the data that runs your workday. Chats, emails, calendar, contacts. If you're all-in on the Microsoft ecosystem, Scout is essentially a proactive assistant that has full context on everything you do and can act on your behalf.
The OpenClaw embrace is strategically sharp. As one observer put it: you only welcome a growing open framework onto your turf when you're confident you can control the ground it stands on. Microsoft is doing what it does best — being a platform company. It gets to ride the agent wave without having to own the entire stack. And unlike Apple, which is unlikely to build a white-glove enterprise experience for OpenClaw, Microsoft is leaning all the way in. The walls of the Microsoft walled garden, it turns out, are actually useful when you want an agent to have broad access without running roughshod over everything.
What Are the New MAI Models Microsoft Trained?
Microsoft announced two in-house foundation models at Build 2025: MAI Code One Flash, a coding model, and MAI Thinking One, a reasoning model. These are the company's first models in their respective categories, and they were positioned heavily around cost-per-token efficiency rather than raw benchmark dominance.
18:50
Scout AI agent Copilot tab demo showing proactive agent actions across Microsoft 365 apps
Watch at 18:50 →
The more interesting comparison isn't with OpenAI or Anthropic — it's with Meta. MAI Thinking One is apparently quite competitive with Meta's recent thinking models, which is notable given how much more attention and hype Meta's AI team has received. Microsoft's MAI team, led by Mustafa Suleyman, hasn't generated the same kind of industry buzz, but the models themselves seem to be performing.
A few other notable details from the model announcements:
- Clean pre-training data: Microsoft made a point of emphasizing that its training data has been sanitized — no New York Times, no Harry Potter, no content from sources that haven't consented. This is a direct pitch to enterprise customers worried about IP liability.
- No distillation from other labs: In a pointed departure from some industry practices, Microsoft stated explicitly that MAI models were not distilled on Anthropic or OpenAI models.
- Fine-tuning options: Microsoft is offering reinforcement learning post-training steps so companies can fine-tune MAI models on their own data. Combined with Azure optimization and predictable token pricing, this is a complete enterprise AI deployment package.
Why Is Microsoft All-In on OpenClaw Open Source AI?
Microsoft's embrace of OpenClaw is a classic platform play. Rather than trying to own a proprietary agent standard, the company is positioning itself as the best place to run the emerging open-source standard. It contributes guardrails, it provides the cloud infrastructure, it sells the enterprise deals — and it lets OpenClaw do the heavy lifting on adoption and ecosystem growth.
The contrast with Apple is instructive. Even though OpenClaw drove a measurable surge in Mac Mini sales, Apple almost certainly won't build a deep enterprise integration for a Microsoft-adjacent open framework. Microsoft, by contrast, has the enterprise motion, the security infrastructure, and the sales team to make OpenClaw genuinely useful at scale. Being a platform has always been Microsoft's superpower — and in the age of agents, that bet looks well-placed.
Is the Cloud the New Hub for Agentic AI Computing?
One of the more philosophical threads running through Microsoft Build 2025 was the question of where agents actually live. Ben Thompson's framing is useful here: thin is in. Because so much AI compute is constrained to data centers, the smartest architecture may be very thin clients connected to very powerful cloud infrastructure.
This is exactly the opposite of how the smartphone era played out, where the phone became the hub of everything and local compute on device was a competitive advantage. For agents, the phone's locked-down nature is a liability. Agents need to move across apps, access enterprise data, and run for minutes or hours in the background. That's cloud work, not on-device work.
Project Solaris is the most literal expression of this thesis: a device whose entire value proposition is being a thin, secure interface to cloud-native agents. Whether or not Project Solaris succeeds as a product, the architectural model it represents — cloud as hub, devices as spokes — seems genuinely compelling for enterprise AI deployments.
Why Did Bots Just Surpass Humans in Internet Traffic?
In what might be the most quietly staggering stat from the week, Cloudflare CEO Matthew Prince reported that bot traffic now accounts for 57.5% of all internet traffic — the first time in the internet's history that bots have surpassed humans. Prince said he expected this milestone to arrive by end of 2027, then revised it to early 2027, but it happened faster than anyone predicted.
The driver is agentic AI. Every time a reasoning model fires up, it's generating hundreds of page requests, tool calls, and API hits. The reasoning traces alone create enormous traffic. As agentic workflows become more common across enterprises, this ratio is only going to climb. The internet, in a very real sense, is increasingly a machine-to-machine network — with humans as the occasional initiators rather than the constant users.
Microsoft Build 2025 was ultimately a conference about that future: one where agents do the work, the cloud is the platform, and the devices we carry are less about consuming content and more about delegating tasks. Whether that vision plays out exactly as Satya Nadella described it is an open question — but the direction is clear, and Microsoft is placing very large bets on being the infrastructure provider for all of it.
