The PCs want you to let them work alone

BarcelonaTaipei and San Francisco have been the scenes this week of a trend change in personal computing. At Computex in Taipei, Jensen Huang announced Nvidia's entry into the PC processor market with the RTX Spark chip, shaking up a sector dominated for decades by Intel, AMD, Qualcomm, and Apple. In San Francisco, Microsoft held the Build 2026 conference at a particularly symbolic moment: exactly two months after formalizing its separation from OpenAI (which has just surpassed 1 billion ChatGPT users), the company wants to demonstrate that it can walk alone in artificial intelligence.

The common thread of the two events has been AI agents: systems that act autonomously, anticipate needs, and execute tasks in the background, increasingly directly on the user's devices, without depending exclusively on the cloud. And the trend will not stop there: next week, Apple will join the race with its WWDC 2026, where it is expected to finally announce a complete and functional version of Apple Intelligence, now supported by Google Gemini models after years of unfulfilled promises with Siri.

Nvidia enters PCs

This week's technical news is the RTX Spark chip, which combines the Grace compute processor (CPU) with the Blackwell graphics chip (GPU), from the same family that powers the most powerful AI systems: 20 CPU cores, 6,144 CUDA cores, and up to one petaflop of AI power, with unified memory of up to 128 GB.

The comparison with Apple's M5, until now the benchmark for local AI, is inevitable. The chip is manufactured by TSMC with 3-nanometer technology in collaboration with MediaTek, and Nvidia has convinced HP, Dell, Lenovo, Asus, MSI, and Gigabyte to adopt it. The first visible fruit is Microsoft's Surface Laptop Ultra: a high-end laptop with a 15-inch mini-LED touchscreen. According to Morgan Stanley, PCs with RTX Spark could start at $1,799 (about 1,550 euros) and go up to $7,000 (about 6,000 euros) in the maximum configuration. By autumn, up to 30 laptop models and 10 desktop models are expected.

At the same Computex, Qualcomm opened fire with CEO Cristiano Amon proclaiming that 2026 is "the year of agents". His thesis is thatsmartphones will cede prominence to agents: "The digital ecosystem no longer resides in the device or the operating system, but in the agents". Qualcomm has also introduced Snapdragon C, a new chip for affordable Windows laptops (from $300, about 260 euros at the current exchange rate), to cover the market segment that Nvidia's RTX Spark does not touch. Demand for tokens—the unit of AI resource consumption—is projected to multiply by 40 by 2030, and Qualcomm aims to be there at all price levels.

There is life at Microsoft without OpenAI

Build 2026 had the pressure to demonstrate that Microsoft can walk alone in AI. The response has been the presentation of seven new proprietary models (the MAI family) that cover reasoning, code generation, images, voice, and audio transcription, available to service creators through Microsoft Foundry. The strategy is to use proprietary models for routine tasks and reserve third-party ones (Anthropic, xAI) for operations that truly require them. Mustafa Suleyman, president of Microsoft AI, has acknowledged that external models are "too expensive". The move also serves them to improve margins on services like GitHub Copilot, which faces growing competition from OpenAI, Anthropic, and Cursor. As a hardware complement, Microsoft has presented the Surface RTX Spark Dev Box, a mini-PC for local AI workloads aimed at developers.

The PC works while you sleep

The most significant novelty of Microsoft Build is Scout, the company's first autonomous agent and the first instance of a new category that Microsoft calls "autopilots." Satya Nadella presented it as "the first agent that acts autonomously, with its own identity and on your behalf." It makes me think of Clippy, the animated assistant that appeared on the Windows screen two decades ago saying "It looks like you're writing a letter, would you like help?" when you least needed it.

Scout is always active and integrated into Teams, Outlook, OneDrive, SharePoint, and the calendar. Unlike the current Copilot, which responds when asked, Scout operates continuously in the background, observes user activity, and executes actions without specific instructions. Among its capabilities are automatic preparation of documentation for meetings, detection of stalled decisions or missed deadlines, and coordination between agents. Over time, Scout builds a semantic model of the user's work life thanks to a layer called Work IQ. It is still restricted access, but it is expected to be opened to everyone before the end of the year.

The longest-reaching bet of Build is Project Solara, a comprehensive platform designed for "agentic devices": a new generation of tools designed from scratch to run AI agents instead of traditional applications. Solara runs on a lightweight operating system not built on Windows, but on AOSP (Android's open-source base), and two prototypes have already been shown with CVS pharmacies, Levi's, and Target supermarkets: a smart medallion and a companion desktop device that rivals Amazon's Alexa+. The idea is that the operating system transcends the device and the cloud simultaneously: a lightweight window to the physical device where the agent manifests itself, while the state and memory persist in Azure (Microsoft's cloud), in a constellation of specialized gadgets that act as a single coordinated system. Qualcomm and MediaTek are already working on the chips that will power these contraptions.

Nvidia, between robots and factories

Beyond the battle for PCs, Nvidia has presented its bet on "physical AI". The most notable is Cosmos 3, the first fully open physical AI "omnimodel", which unifies visual reasoning, simulated world generation, and physical action generation in a single architecture for training robots and autonomous vehicles in photorealistic simulation environments. Nvidia has also presented the Isaac GR00T reference humanoid robot, which combines Unitree's H2 Plus chassis with Sharpa Wave's tactile hands and the Jetson Thor electronic module. The philosophy mirrors that of CUDA software: building an open ecosystem and first attracting researchers (Stanford, ETH Zurich, UC San Diego) to ensure long-term adoption.

All this is relevant for the common user because there is an underlying shift. During the early years of generative AI, the dominant model has been the cloud model: the user makes requests via a conversational bot to a remote server and waits for the response. But local AI execution, with models running directly on the device without transmitting data to any server, is already a commercial reality. It is in this arena that the RTX Spark, Qualcomm's Snapdragon X2, and Apple's M5 compete. RTX Spark is the most ambitious proposal seen so far in a consumer laptop, and it makes it possible to locally execute models that until now depended on services like ChatGPT.

However, neither Computex nor Build 2026 have resolved the fundamental questions: who pays for the energy consumption of agents, who controls the data, how is the autonomy of systems that act "on your behalf" without asking for your permission for each action governed. But it seems clear that the personal computer we know, which waits for the user to tell it what to do, is becoming a thing of the past.