Perplexity's hybrid AI orchestrator keeps sensitive data local, cuts cloud costs

Perplexity CEO Aravind Srinivas took the stage at Computex 2026 in Taipei on June 2 to announce what the company calls the first hybrid local-server inference orchestrator — a system that automatically decides, task by task, whether AI work runs on a user's own device or gets routed to powerful cloud models. Intel CEO Lip-Bu Tan joined Srinivas for the announcement, with the demo running on Intel's Core Ultra Series 3 processors. The feature, branded 'hybrid agentic inference,' will arrive in July as an update to Perplexity Computer, the company's Mac-native always-on agent product that launched in March at $200 per month. Rather than forcing users to choose between local or cloud processing upfront, the orchestrator splits and coordinates tasks automatically — keeping sensitive data like financial records and health information on-device while offloading computationally intensive reasoning to frontier cloud models. At its core, the system runs a lightweight model on the user's device that evaluates each incoming task for sensitivity and complexity. Simple operations — document summarization, text formatting, lightweight classification — execute locally. Tasks requiring deeper reasoning get routed to Perplexity's cloud-based frontier models. The key design decision is that splitting happens at the sub-task level, not the session level. A single workflow involving a private financial document, for example, might keep the raw data local while sending an anonymized summary to cloud models for deeper analysis.

The hybrid approach serves a dual purpose: privacy and cost reduction. By offloading inference to user hardware, Perplexity reduces its own cloud compute bills — a significant expense as the company scales. Perplexity's revenue grew from $100 million to $500 million while headcount increased only 34%, suggesting the company is already focused on operational efficiency. Srinivas emphasized that the orchestration layer is chip-agnostic. The Computex demo ran on Intel Core Ultra Series 3 processors, but Perplexity confirmed the same system works on Nvidia's RTX Spark platform and other local silicon. The announcement positions Perplexity against a growing field of companies — including Apple, Google, and Microsoft — racing to blend on-device and cloud AI processing. But Perplexity frames its approach as uniquely automatic: a compact local model acts as a traffic cop, making real-time decisions about where each piece of a workflow should execute.