The pitch from Jensen Huang was characteristically grand. “For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask, and the PC does the work,” Nvidia’s chief executive said at Computex, announcing a chip designed to bring AI agents to personal computers. The Wall Street Journal reported the hardware push, which Nvidia frames as the shift from the PC as a tool to the PC as an assistant. RTX Spark, unveiled at the company’s GTC Taipei keynote, is Nvidia’s first serious attempt to put its AI platform inside consumer and professional desktops.
The chip itself is a superchip: a 20-core Arm-based Grace CPU paired with a Blackwell GPU with up to 6,144 CUDA cores, connected through Nvidia’s NVLink-C2C interconnect to as much as 128 gigabytes of unified memory. Nvidia says the design delivers up to one petaflop of AI compute, enough to run models with up to 120 billion parameters locally, and that the graphics performance rivals a desktop GeForce RTX 5070 in an 80-watt envelope. The first systems arrive this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and GIGABYTE to follow.
The software story is where the ambition shows. Nvidia and Microsoft are jointly building a layer for local AI agents inside Windows, combining new Windows security primitives that cover identity, containment, policy and end-to-end security for native agents, with Nvidia’s OpenShell, a runtime designed to let agents run safely on-device. OpenShell adds policy controls that let users define what an agent can and cannot do, a recognition that agents, which act on their own, need guardrails that ordinary apps never required. Adobe is optimizing Photoshop and Premiere for the platform, and Nvidia claims a twofold increase in AI and graphics performance for creative workloads, a figure independent testing has not yet confirmed.
The strategic logic is simple: agents are moving from the cloud to the edge. Running an AI agent in the cloud means sending data to a data center, paying for inference, and living with latency and connectivity dependence. Running it locally means privacy, speed and predictable costs. Enterprises are increasingly interested in agents that handle sensitive documents, customer data or code without ever leaving the building, and Nvidia’s bet is that the hardware platform for that shift is worth building, and that it will sell a lot of GPUs in the process.
The market reaction has been muted compared with Nvidia’s data center announcements, and analysts said the agent PC is a longer-term play. The desktop AI market is real but young, and the installed base of PCs with enough memory to run frontier-class models locally is tiny. Nvidia’s answer is the DGX Station line and its Agent Toolkit, which the company says lets developers deploy and run AI agents locally in about 30 minutes through a three-step process, compressing setup work that once took days or weeks. The toolkit, paired with the GB300 Blackwell Ultra GPU, targets the developers and enterprises that want to build agent applications without renting a cloud cluster.
The competitive picture is complicated. Qualcomm, Intel and AMD already sell AI PC chips, and Microsoft’s Copilot Plus initiative has pushed neural processing units into mainstream laptops. Apple’s silicon runs local models well and has set the standard for unified memory architectures. Nvidia’s advantages are the CUDA ecosystem, which dominates AI development, and the company’s ability to pair hardware with software that developers already use. The question is whether those advantages extend from data centers, where Nvidia is nearly uncontested, to desktops, where it is entering late against entrenched rivals.
The security questions are also unresolved. An agent that can read files, send messages and act on the web is a powerful tool and a powerful target, and running agents locally changes the threat model in both directions. Local execution keeps sensitive data off cloud servers, but it also moves agents outside the visibility of enterprise security tools that were built to monitor cloud workloads. The Microsoft collaboration is an attempt to answer that, with operating-system-level controls that treat agents as first-class citizens with their own identity and permissions. Whether those controls hold up against real attackers is a question for the security community to test once the systems ship.
There is also a pricing question. Nvidia has not disclosed RTX Spark system pricing, but the component costs, a 20-core CPU, a 6,144-core GPU and up to 128 gigabytes of unified memory, suggest premium positioning. The fall launch will tell whether buyers pay a premium for local agents or treat them as a nice-to-have. Huang’s framing suggests Nvidia believes the shift is inevitable: if software is becoming agents, and agents need local compute, then the PC market will eventually be rebuilt around AI silicon. That is the same argument the company made about data centers in 2023, and it was right then.
For Nvidia, the agent PC is part of a larger strategy of selling AI infrastructure everywhere, from cloud to workstation to edge. The company’s data center business has made it the most valuable chip maker in history, and the desktop is the next frontier of compute demand. Analysts said the RTX Spark launch is less about immediate revenue than about establishing a beachhead, getting Nvidia silicon and software into the machines where the next generation of AI products will be built. If agents do move to the desktop, the company that owns the silicon, the runtime and the developer tools will own the platform. Nvidia intends to be that company.
