(Credit: Microsoft)
SAN FRANCISCO—What stood out most for me at this year’s Microsoft Build conference was the emphasis Microsoft put on how anyone can "participate fully" in the new era of AI and agentic computing.
As Build is a developers conference, they were the intended audience. However, there was a similar message for enterprises. CEO Satya Nadella stressed that organizations could use their own data to fine-tune models and create and manage their own agent ecosystems, while keeping costs in check. The company even applied that concept to individuals in the communities where Microsoft wants to build data centers. Nadella said Microsoft has to prove it won’t raise electricity costs or use a significant amount of water, while contributing to the tax base, and helping local organizations.
There were lots of Windows-specific announcements, including new ways to get a clean desktop and move the taskbar, as well as a new Intelligent Terminal designed for developers who would put the regular terminal in one window and an agent in the other. Perhaps the most important of these announcements is the ability to run agents in containers that can be sandboxed and have their own permissions, all controlled by the individual developer and the organization. These Microsoft Execution Containers (MXC) isolate agents so they can't do damage to other systems or resources (such as a rogue agent accidentally deleting a database).
Within these containers, you can run agents, including an OpenClaw. (Running OpenClaw on Windows was a big theme throughout the show.) It's a powerful tool that can do many things on your behalf, but organizations have been reluctant to allow users to run it on their machines because of the potential for mischief. The new container framework should mitigate many of those concerns. The company also described other ways of creating long-running agents that work on your behalf as "autopilots."
'This Is a New Era in AI'
Microsoft AI Head Mustafa Suleyman also announced seven new Microsoft AI models, including a general model, the company's first reasoning model, and new models for creating images, transcriptions, speech, and code. Microsoft made a big deal about having a "clean lineage" with lots of transparency about how they were trained.
Interestingly, the company didn't claim its new models are the best, just the most cost-effective for many tasks. "This is a new era in AI...that you control on your terms,” Suleyman said.
This concept of grounding AI within the organization is something Microsoft stressed at its Ignite conference last fall, where it introduced WorkIQ, a layer that includes information stored in Microsoft systems such as email, Teams, OneNote, and SharePoint. At Build, it announced a new component, WebIQ, which it touted as the fastest way to get real-time web data. Combining this with information from the AI tools and agents (Foundry IQ) and its data warehouse (Fabric IQ), users can create a comprehensive context for AI systems and agents. This seems to be a major way Microsoft wants to differentiate itself from other agentic platforms.
More to the point, Nadella described how organizations could fine-tune these models using their own internal data to create a "hill-climbing" AI tailored to their knowledge and ways of working. I doubt that smaller organizations or individuals will do much of this, but I can see where it would be attractive to larger enterprises.
Microsoft devoted a lot of time and floor space to its new hardware, specifically a Surface Laptop Ultra and an RTX Spark Dev Box, both running a new Nvidia RTX Spark processor that the company says can support up to 128GB of shared memory and run up to 120-billion-parameter local models. These are due in the fall, and look very powerful, especially for AI developers. But of course, the changes to Windows and the development tools should work on any machine, and Microsoft continues to push the use of Windows 365 cloud instances for development as well.
The big benefit of running local models is that they incur no additional costs because they use the device's processors. Nadella called this "unmetered intelligence," and given the costs many organizations have run into with cloud-based tools, I can see how this could be very attractive.
Of course, it's not just local PCs that are changing in the move to "agentic computing." Nadella discussed how the move to agents was affecting the entire computing stack, from infrastructure to underlying models and tools to security.
In another session, Microsoft EVP for Cloud and AI Scott Guthrie described the significant changes in building data centers today, from new networking approaches to more automated, managed services—all necessary because new data centers are much larger and being built at a faster pace than ever. Nadella said Microsoft has added more data center capacity in the last 18 months than it did in the first 10 years of Azure.
Nadella ended his keynote by reminding the audience that, in general, there are two possible stories we can tell about the move to AI and agentic computing. In one, technology concentrates power and reduces human agency; in the other, we use the new technology to unlock more opportunities for everyone. Making the second alternative a reality is a north star for Microsoft.
Taken together, these announcements should make AI development and deployment easier for developers and safer for organizations. All this is crucial, and we've heard similar concepts from Google and AWS in recent months. Still, I would argue there's still a lot more to do. Figuring out how we'll implement all this so we get real value is going to be the biggest issue for organizations for at least the rest of the decade.