Perplexity AI

Perplexity AI is an AI-powered answer engine founded in 2022 by Aravind Srinivas (former OpenAI and Google researcher), Denis Yarats, Johnny Ho, and Andy Konwinski. Valued at over $9 billion, Perplexity has emerged as the leading AI alternative to traditional search engines.

Perplexity combines large language models with real-time web search to provide direct, cited answers to user questions rather than lists of links. The platform searches the web, reads and synthesizes multiple sources, and generates comprehensive responses with inline citations — allowing users to verify every claim. Perplexity Pro offers access to more powerful models and advanced features like file analysis, image generation, and deeper research capabilities.

Perplexity has expanded its product line to include Perplexity Assistant (AI agent for completing tasks), Comet (AI-powered web browser), and enterprise solutions. The company processes hundreds of millions of queries and has attracted users who prefer AI-synthesized answers over traditional search results. Perplexity's rapid growth has drawn attention from Google and other search incumbents, validating the thesis that AI will fundamentally transform how people find and consume information online.

Perplexity's growth strategy has centered on owned surfaces (perplexity.ai, the Comet browser, Perplexity Assistant) plus enterprise and API channels rather than embeds in third-party consumer apps. A $400 million partnership with Snap to integrate Perplexity's answer engine into Snapchat's "Chat" interface was tested with a select user group and ended after the two companies failed to converge on a broader rollout — a useful signal about the cost of grafting AI search onto incumbent consumer surfaces.

On August 25, 2026 Perplexity released Portable Computer with NVIDIA, moving the entire agent loop onto local hardware with no credit cost for work done on-device. The design is unusually conservative for an agent product: the orchestrator running the loop is deterministic code rather than a second model, tool calls execute in an operating-system-level sandbox, and the harness fails closed — disabling itself if the sandbox is unavailable rather than running unsandboxed. Escalation to a cloud model is permission-gated and shows the user exactly what would leave the machine, with the cloud model returning text guidance only. It requires an NVIDIA DGX Spark or an RTX card with at least 24 gigabytes of video memory.