Perplexity has expanded from search, browsers and cloud-based AI agents into local AI services, unveiling a new tool called Portable Computer. The move comes amid reports that Nvidia is considering a $30 billion (£23.6 billion) investment in Perplexity, and the new product runs on Nvidia's DGX Spark hardware, prompting speculation about the timing and heavy branding of the announcement. Perplexity frames the shift as a way to cut AI running costs and address privacy concerns, as more users turn to locally hosted open-weight models to avoid expensive cloud API bills.

Portable Computer combines an agent harness, an orchestrator and local AI models running on a DGX Spark workstation, with the option to draw on cloud inference when needed, building on Perplexity's existing hybrid inference orchestrator. The company highlighted small, efficient models such as Nvidia's Nemotron 3.5 Lightning, alongside Qwen 3.6 and Qwen 3.8, saying they can now handle complex agentic workflows. Perplexity published benchmark comparisons against rival agent harnesses Pi and Hermes, claiming Portable Computer matched or beat them on accuracy while being fastest on the BrowseComp and ParseBench-100 tests and using fewer tokens throughout, though it noted that competing harnesses are currently free.