Perplexity is trying to move part of its AI workload back onto the user's computer. At Computex in Taipei on June 2, CEO Aravind Srinivas appeared with Intel CEO Lip-Bu Tan to show what the company calls hybrid agentic inference.
The idea is to run a lightweight local model as a dispatcher. Simple tasks such as summarizing a local document, reformatting text or classifying small items can stay on the PC; harder reasoning goes to the cloud; sensitive records such as finance or health data remain local unless the user approves sending them out.
The first version is planned for July inside Perplexity Computer, beginning with the Windows PC app. The demo used Intel Core Ultra Series 3 hardware, but Perplexity said the approach is chip-agnostic and can also work with Nvidia processors.
The cost logic is clear. Srinivas warned that companies do not want all compute centralized in servers and said some are spending half a billion dollars a month. For Perplexity, splitting work between local PCs and cloud models could reduce inference pressure while giving enterprises a stronger data-governance story.
Sources: VentureBeat, Decrypt, 9to5Mac, CocoLoop.