Perplexity rolled out Portable Computer on Tuesday, a local AI agent running on Nvidia DGX Spark hardware that keeps private data on-device and eliminates per-token inference costs. Developed in partnership with Nvidia, the service uses Qwen 3.8 27B or PPLX 27B models to process workflows locally while routing advanced tasks to the cloud only when explicitly authorized.
Architecture and Local Control Plane Mechanics
The system shifts the operational paradigm away from purely cloud-metered API dependencies. Running on the Nvidia DGX Spark, Portable Computer executes its orchestrator, planner, tool router, scheduler, durable task queue, and local search index entirely on-device. Aman Mahapatra, chief strategy officer for Tribeca Softtech, noted that while running models locally has been standard for two years, executing the entire agentic control plane locally changes how decisions are made. The model itself determines whether a task requires cloud escalation based on post-trained boundaries.
https://x.com/perplexity_ai/status/2092268362386780270
Under the hood, the offering currently requires Linux as its underlying operating system, with Windows support slated for release soon. Users interact with applications via built-in connectors linking the local environment to Google Drive, Gmail, Slack, and GitHub. According to Perplexity, this setup allows confidential data—such as term sheet details—remain local while allowing the orchestrator to fetch current market comparisons or precedent deals from the cloud.
Hardware Demands and Enterprise Economics
Despite the zero-token-cost promise for local execution, industry analysts point out that the hardware barrier to entry is substantial. Flavio Villanustre, CISO for the LexisNexis Risk Solutions Group, emphasized that the setup requires significant initial capital investment. Specifically, the architecture demands a local GPU with a minimum of 24GB of VRAM.
Gartner VP Analyst Nader Henein added that while select models can run on high-end laptops with off-the-shelf GPUs, broader enterprise adoption depends heavily on the final pricing structure of the hardware and software package. For builders and infrastructure teams evaluating edge AI, balancing upfront hardware expenses against recurring cloud API fees remains a critical financial calculation.
Security, Egress Governance, and the Consent Dilemma
Enterprise security professionals have raised questions regarding data egress governance and how local-versus-cloud decisions are enforced. Justin Greis, CEO of consulting firm Acceligence, warned that local-first must not be confused with local-only. Because autonomous agents operate across complex file structures and application connectors, users might routinely approve escalation prompts without fully grasping the scope of data crossing the network boundary.
Mike Wilkes, enterprise CISO at Aikido Security, noted that while IT administrators could technically block all external network access, such a lockdown would neutralize the utility of hybrid workflows where connectivity is essential. For instance, a proprietary trading firm might process sensitive positions locally while still requiring real-time market data or SEC feeds.
Expanding on these structural risks, Aman Mahapatra argued that relying on user consent via a permission prompt fails in adversarial scenarios. Because the gate depends on a probabilistic model classifying sensitive content correctly, and on a user inspecting a complex payload, sophisticated prompt injection tactics could theoretically trigger unauthorized data migration. Mahapatra stated that what enterprises require for compliance is network-layer enforcement—such as a mandatory egress proxy, deterministic classification rules, and immutable logging—rather than application-layer prompts.
Perplexity Response and Platform Guardrails
In response to security inquiries, Perplexity Communication Manager Beejoli Shah pushed back against the possibility of unvoted data migration. Shah stated that content inside a local document cannot trigger an escalation independently and cannot override product controls. Moving a task to the cloud requires an explicit, per-action approval from the user, alongside toggling the app out of its default local-only mode.
Furthermore, Perplexity confirmed that users must manually activate the “allow advisor escalation” setting in the application preferences. Without this toggle engaged, no computational work proceeds to the cloud. The platform also restricts escalation to a single instance per request, preventing automatic propagation across the remainder of a task or into future sessions.
