Running the agent takes Nvidia's DGX Spark, or a Linux PC with an Nvidia RTX card carrying at least 24GB of video memory. In practice that means an RTX 3090 or newer.
Perplexity released Portable Computer, a version of its agentic Computer platform built with Nvidia that runs on a desktop GPU instead of in the cloud. Work the agent finishes there is not billed by the token.
An agent working for hours through a folder of documents, checking its own output and rerunning analyses, bills every one of those tokens in the cloud. On hardware already paid for, the same run costs next to nothing at the margin. Nader Khalil, Nvidia's director of developer technology, argues that agents should be kept running and made to burn through as many tokens as they can.
In one briefing demonstration the agent played an individual investor working through a folder of 1099s and investment paperwork. A 27-billion-parameter Qwen model kept the GPU of a DGX Spark fully occupied, going document by document and flagging fees charged needlessly. Nate Kupp, Perplexity's vice president for enterprise and infrastructure, said the interface's running count of cloud credits stood at zero.
The agent can hand a step up to a frontier model in the cloud when the local model runs out of road. Perplexity says every task opens on the user's own hardware, and that the software shows what would leave the device and asks for approval before it goes.
Perplexity has put numbers on what staying local costs, all from evaluations it ran itself. On Terminal Bench 2.1, a coding benchmark, the Qwen model working entirely on the device reached 59.6% at no marginal cost, against 82.4% for a frontier model at $0.65 a task. Calling a Claude Opus 5 advisor from the local run landed between them, at 73.0% and about $0.415.
Portable Computer is open only to Perplexity's paying subscribers. It runs on Linux alone for now, with Windows support due in September; Apple silicon, a favorite of local AI enthusiasts, does not appear on the roadmap at all. For Nvidia the launch gives the DGX Spark something to do. By the admission of people at the briefing, the machine has been easier to buy than to run.