“Portable Computer” reaches Windows: the AI agent leaves the cloud for Nvidia cards

Perplexity has launched its on-premise version of the multi-task agent under the name “Portable Computer” for Windows users, relying on Nvidia graphics processors from the GeForce RTX and RTX PRO families, in a move that shifts execution of software agents from exclusive cloud reliance to direct desktop hardware. This update enables running specialized local models to perform planning, financial and technical execution tasks without exporting sensitive files outside the device, and exempts locally completed operations from consuming cloud-service account credits.

The system relies on a custom local model, such as “Qwen 3.8 27B”, which was subsequently fine-tuned to align with Perplexity’s agent architecture and Nvidia graphics processors. The platform integrates internal execution tools, including an embedded browser and a secure sandbox environment called SPACE, allowing the agent to read documents and intercept data via software connectors linked to Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub, without requiring the user to build a complex software stack or manually configure runtime platforms.

When the agent requires advanced reasoning capabilities beyond the local model’s capacity, the system asks the user for permission before routing the task to the cloud.This hybrid design ensures that sensitive documents remain under the organization’s direct control, while providing an optional pathway to cloud systems for complex research needs. In practice, the mechanism appears in automated review of pull-request submissions on GitHub and automatic updates of technical documentation, in analysis of corporate digital activation logs with export of rollback summaries to Slack channels, and in examination of two years of tax records and investment portfolios to pinpoint recurring fees with page- and line-level precision, all without any document leaving the premises for external servers.

Running the tool requires a graphics card with at least 24 GB of video RAM on compatible Windows machines, an expansion that follows earlier support for Linux systems and DGX Spark workstations, with expectations of DGX Station support soon. This hardware prerequisite limits the feature to professional-grade devices and makes the memory requirement the primary threshold for replacing cloud services with desktop processing.

For teams in the Gulf states and Egypt, this shift changes how they deal with data-protection and banking-confidentiality constraints. Financial compliance, legal audit departments and law firms in Riyadh, Abu Dhabi and Cairo face regulatory restrictions that prohibit sending contract drafts or tax filings through third-party cloud chat interfaces, in compliance with privacy and data-governance frameworks. Providing a local agent that inspects files within the work environment opens the door to automating these complex tasks without exposing the organization to regulatory risk, but it also forces a change in spending structure, moving costs from recurring cloud subscriptions to direct investment in upgrading desktop workstations with advanced graphics cards.

The move toward local models presents technology leaders with a new equation that balances hardware cost against data security.If you lead a technical or analytical team, the equation today requires you to assess the volume of records that cannot be moved to the cloud against the cost of equipping inspection workstations with cards that have enough memory to run these models locally. The practical step begins with cataloguing workflows that demand absolute confidentiality and testing whether desktop processing can meet continuous-analysis requirements without draining cloud budgets.