Perplexity and Nvidia launch a fully local AI agent

Perplexity and Nvidia announced Portable Computer on August 25, a version of Perplexity’s existing Computer agent, first released in February for multi-step tasks like building web apps, that now runs entirely on local hardware instead of Perplexity’s cloud. Because inference happens on the user’s own machine, private data stays on-device and local runs don’t draw down Perplexity subscription credits. Perplexity is offering it as a free add-on to existing subscriptions rather than a separate paid product.

At launch, Portable Computer only works on Nvidia’s DGX Spark or a Linux workstation with an Nvidia RTX GPU that has at least 24GB of VRAM, making this a Linux-first, high-end-hardware release rather than a laptop-friendly one. Perplexity and Nvidia are positioning it as a way to keep sensitive workflows off the cloud entirely while still getting agentic behavior, multi-step task execution rather than single-turn chat.

What it means for you

This is squarely a hardware-gated release: if you don’t already have a DGX Spark or a 24GB-plus RTX card sitting idle, there’s nothing to try yet. For operators who do have that hardware, or are weighing Apple’s local-AI Mac Studio economics against an Nvidia-based alternative, the zero-token-cost pitch is worth testing against your actual workload before assuming it beats a cloud agent on cost, since local compute still has real electricity and hardware-amortization costs that a per-token bill makes visible and a “free” local run does not.