Perplexity Ships Portable Computer on NVIDIA DGX Spark: Local Harness, OS-Enforced Sandbox, and Zero Per-Token Cost for Local Steps

Perplexity has released Portable Computer, a system that runs AI agent software directly on local NVIDIA hardware rather than relying on cloud servers. The system includes a local model, inference engine, orchestrator, and security sandbox, with tasks processed locally at no per-token cost, though steps requiring live web access or advanced reasoning require explicit user approval before sending to cloud models. This matters for organizations in finance, legal, healthcare, and government where data residency rules or confidentiality requirements prevent sending information to cloud services. The system requires significant hardware investment, starting with either a GB10-class NVIDIA box or an RTX GPU with 24 GB of VRAM, making it viable only for enterprises and well-funded startups rather than general small businesses.