Perplexity Wants Your Laptop to Do Part of the AI Work—So It Doesn’t Have To

Perplexity CEO Aravind Srinivas took the stage Computerex 2026 in Taipei on June 2 with Intel CEO Lip-Bu Tan to announce what the company is calling the first hybrid inference orchestrator for on-premises servers. Coming to Perplexity Computer in July, the system automatically decides which parts of an AI job to run on your machine and which parts to route to more powerful models in the cloud, without asking you to make a choice.

“Today we’re announcing the next step for Personal Computer: the first hybrid inference orchestrator for on-premises servers,” Perplexity announced. “It decides what work should be done on your device and what work should go to cloud agents, automatically routing each part of a job to the right place”

“The proper goal for an AI system is to deliver the highest token value per watt for each user,” Perplexity wrote in the official announcement. Three competing factors make that difficult: accuracy requires the most capable models, privacy requires that some data never leave your machine, and cost requires that you don’t spend the computing resources of a frontier model on a task that a smaller model can handle.

The solution, which Perplexity calls “hybrid agentic inference,” addresses all three at once. A compact model runs locally on your device and acts as a traffic cop: it figures out what information is sensitive enough to stay local and which tasks need the full power of a cloud-based boundary model.

“Hybrid agentic inference is intended for work that contains sensitive data but needs powerful AI, such as financial data, health information, and personal files,” the company explains. “The compact model runs locally on your device to determine when sensitive data should also be kept locally. Meanwhile, work that requires the full capacity of a frontier model runs on the server.”

Should you take it into account?

Inference (the process of running a trained AI model to generate a response) is the computation that happens every time you send a prompt to a chatbot. Right now, almost everything happens on remote servers owned by AI companies. That means your financial documents, health questions and private notes go to someone else’s computer before you get an answer.

This is why you see ‘Auto’ modes or ‘low thinking’ modes on your chatbot. AI companies will always try to force users to route interactions in the cheapest way for them.

Srinivas has been direct about this. In a Bloomberg Television interview at Computex, he said the quiet part out loud: “You don’t want all your computing power to be centralized on servers and run through the biggest models. Some people are spending half a billion dollars a month. What you really want is an efficient value per watt per user.” Offloading the inference work to user hardware reduces these bills – for Perplexity.

Local inference is best for those companies because it saves a lot of the cost, but it has an important benefit for AI users: it keeps that data on your machine. The trade-off has always been power: smaller models that run locally are less capable than the large models that live in data centers.

Perplexity’s orchestrator tries to get both. Simple tasks (summarizing a document you’ve already written, formatting text, lightweight classification) are performed locally. Complex reasoning is routed to the cloud, ideally without the sensitive parts of your task attached. The company says this happens automatically, mid-task, invisible to the user. Whether the routing is as reliable in practice as it sounds in a Computex demo is a question that the rollout in July will answer.

One clarification worth making: this isn’t Perplexity giving away an open-source on-premises model that you control. The local component is a compact model that Perplexity uses as part of its app. The cloud component still runs via Perplexity’s servers. Users who want a completely offline, self-hosted setup, as projects like MiniCPM5-1B offer, won’t find that here.

The figures reflect that framework context. Perplexity’s revenue grown from $100 million to $500 million, while increasing headcount by only 34%, Srinivas announced in April. A company that directs queries to models it does not train has strong incentives to keep computational costs as low as possible. Shifting some of the inference burden to users’ devices (billions of PCs already in circulation) is an efficient way to do that. The privacy field is real, but it easily ties in with the financial one.

Who else does this

Every major player in AI is currently pursuing on-device or hybrid inference. Apple Intelligence performs the most sensitive processing locally on M-series chips. Microsoft’s Foundry Local became generally available in April 2026, enabling full AI inference on Windows, macOS, and Linux without cloud dependence.

Nvidia announced RTX Spark at the same Computex where Perplexity made its announcement, targeting local LLM inference on laptops and desktops. Google’s approach, such as Decoding reportedis more controversial: Chrome quietly installed a 4GB Gemini Nano model without user consent, and the “AI Mode” button that most users actually see doesn’t even use it.

The differentiation of Perplexity is the orchestration layer. Instead of asking users in advance to choose local or in the cloud, the system decides per task in real time. Srinivas said the approach is “chip agnostic”: the Computex demo ran on Intel Core Ultra Series 3, but Nvidia processors are also supported. The feature is currently exclusive to the Perplexity for Windows PC app, with a wider rollout timeline not yet confirmed.