Hands On With Hybrid Agentic Inference: How Intel and Perplexity Kept 'Project Falco' Confidential at Computex

At Computex 2026, Intel CEO Lip-Bu Tan and Perplexity CEO Aravind Srinivas took the stage to answer a critical enterprise question: can an AI be both powerful and truly private? Their answer was a live demonstration of hybrid agentic inference, a new computing architecture designed to keep confidential data off the cloud entirely by running sensitive AI workloads on a local device .

The demo simulated a private equity analyst tasked with evaluating a confidential deal, codenamed "Project Falcon." The analyst's files, which included non-disclosure agreements, financial models, and internal transcripts, were processed by an AI model running directly on a laptop powered by an Intel Core Ultra Series 3 processor, the first chip built on Intel's advanced 18A process technology . The Perplexity Computer, an AI operating system that orchestrates a team of up to 20 different AI models, first analyzed the documents on-device to determine which information was sensitive and could not be shared, and which public research tasks could be safely outsourced to the cloud . The result is a system that uses the cloud for heavy lifting—like web research—while ensuring truly private data never leaves the local machine .

The Architecture of Hybrid Agentic Inference

The core innovation lies in the real-time decision-making layer of the Perplexity Computer OS. It acts as an "air-traffic controller" for AI tasks, dynamically routing workloads based on data sensitivity, latency requirements, and cost .

• Local Privacy Vault: Smaller, specialized models run directly on the Core Ultra Series 3's combined CPU, GPU, and NPU to analyze documents, identify personally identifiable information (PII) and trade secrets, and handle edits to local files. This ensures that confidential materials are never uploaded to a server for processing .
• Cloud Orchestration: Once local tasks are complete, the system dispatches public-web queries and complex reasoning tasks to more powerful models on Perplexity's cloud servers. This offloads computational burden and reduces inference costs by an estimated 30-50% compared to running everything in the cloud .
• Token Value Per Watt: Intel and Perplexity framed the system as a way to maximize "token value per watt," using efficient local silicon for simple tasks and reserving expensive data-center compute for jobs that truly require it .

This approach is not purely theoretical. The Computex demo showed this capability working on a Windows PC with an Intel Core Ultra Series 3, though the feature remains exclusive to that specific Intel-powered Windows version of the Perplexity app for now .

How It Fits into Intel's Broader Computex 2026 Strategy

The Perplexity demo was the client-side proof point for a much larger Intel narrative: that the era of agentic AI is driving a "CPU renaissance," where the central processor becomes the critical orchestration layer for vast networks of AI agents .

Xeon 6+ and the 150,000-Agent Rack

For the data center, Intel announced the Xeon 6+ processor (Clearwater Forest), its first server CPU built on the Intel 18A process . A single 32U liquid-cooled rack packed with these processors can deliver over 36,000 cores and is designed to support up to 150,000 AI agents—the highest agentic AI compute density Intel has claimed . The processors are built for the orchestration, concurrency, and data-movement demands of agentic workloads, reclaiming power and rack capacity in the process .

Client and Edge Momentum

Intel also provided key updates on the commercial progress of its 18A technology, announcing:

• Over 300 design wins for the Core Ultra Series 3, spanning nearly 400 PC designs across OEMs .
• Over 130 edge AI designs in deployment or development, supported by 4,000+ edge ecosystem partners .

A Heterogeneous Compute Future

CEO Lip-Bu Tan described a strategic vision framed around four multi-trillion-dollar compute markets: personal computers, edge and physical AI, data centers, and digital intelligence centers . He emphasized that future AI infrastructure would not be monolithic but heterogeneous—blending CPUs, GPUs, ASICs, and custom chips to handle specific workloads most efficiently . In a separate announcement, Intel partnered with SambaNova and Foxconn on a new "disaggregated inference" rackscale architecture, further demonstrating this chip-agnostic vision .

Perplexity's Own Hybrid Play for Mac Users

The Computex demo was an Intel partnership, but the underlying hybrid philosophy was already extending to other platforms. By May 2026, Perplexity had independently rolled out its Personal Computer agent to all Mac users on Pro and Enterprise plans .

The Mac version creates a hybrid local-and-cloud runtime. A Mac or Mac mini can run the agent continuously, giving it sandboxed access to local files, native apps like Mail and Calendar, and the Comet web browser, while still connecting to Perplexity's secure servers for heavy orchestration and access to frontier models . While it lacks the Intel-specific on-device AI inference at the hardware level, it reflects the same product strategy: make the AI more capable by making it more personal, without compromising security .