Microsoft builds agentic Windows 11 around free local AI and a $2,599 Surface laptop

The Verge's reviewer switches off the Bing results in Windows Search ^[14]. The change at Microsoft's Windows and Surface event that suited that habit was a redesigned Search, much faster than before and showing results inline so it works more like a launcher ^[15]. "For the rest of us, Microsoft had very little to show," the reviewer wrote ^[16].

Satya Nadella, Microsoft's chief executive, spent part of his talk on the security, identity and management of agents, close enough to a business pitch that The Verge remarked on it ^[3]. He also said, "I think today is the day we'll start talking and thinking about the frontier ecosystem taking maybe slightly different contours" ^[4]. The flagship hardware has a narrower audience. "We built it for developers and creators that push the limit of performance and want portability with exceptional craft," Pavan Davuluri, the Windows chief, said of the Surface Laptop Ultra ^[7].

Hybrid intelligence is a cost argument. Windows takes advantage of free local models in place of expensive cloud ones ^[2], and Davuluri described the goal as intelligence "that moves seamlessly between local and cloud" ^[11]. The local route is paid for up front. The top of the RTX Spark range costs $4,400 more than the entry Surface Laptop Ultra ^[17]. The Verge's case for paying is specific. "For many AI builders it might just be worth it, though, especially if you can save thousands in token costs," the reviewer wrote ^[18]. The account of the event does not include token prices or the admin controls that decide when a task leaves the device. Any payback figure for a seat has to come from that team's own cloud bill.

Copilot Plus PCs are the precedent. Here's what buyers were told users would do with an NPU: change how they use everyday AI on a laptop ^[12]. Here's what users actually do two years later: open one of the handful of photo, video and productivity tools that use it ^[12]. Recall, the feature meant to show the NPU off, ended up irrelevant after privacy concerns and a delayed launch, according to The Verge ^[13]. The reviewer allows that "Nvidia's presence and GPU dominance may well make RTX Spark a bigger success story than Copilot Plus PCs" ^[19].

I'd sort seats on two facts: whether the machine can run the model locally, and whether the user already runs up a cloud token bill you can measure. The developer with a heavy bill on an RTX Spark laptop is the case Microsoft demoed ^[5], and the only one where the hardware price can be set against real spend. Heavy users on ordinary hardware stay on cloud and get budgeted as tokens. Put a light user on local-capable hardware and you have the Copilot Plus PC case again, with paid-for silicon waiting on software. Light users on ordinary hardware get the new Search, and their AI cost does not move.

My recommendation is to buy RTX Spark machines only for that first group and to move people in as their bills grow. The tradeoff is a fleet with two AI cost models, a few seats on local inference and the rest on metered cloud, plus a refresh decision tied to token prices the IT team does not set. The test is one comparison per seat. Carry last quarter's token spend across the laptop's service life and set it against the extra cost of the RTX Spark machine over the laptop that seat would otherwise get. If the token figure is smaller, the seat stays on cloud.