Perplexity Computer landed on Windows this week. The pitch: sensitive files like financial records or health data get processed entirely on your machine, while only the heavy reasoning tasks get routed to the cloud. Here's the thing though — the company making that promise is currently facing 9 copyright and trademark lawsuits.
What Actually Shipped This Week
Perplexity's desktop agent has been building up to this. It started as a cloud-only "Computer" on February 25, went Mac-only as "Personal Computer" on March 13, then opened up to Pro tier ($20/month) subscribers on May 7. And on June 3, Windows finally got its turn.
Just the day before, at Computex 2026 in Taiwan, CEO Aravind Srinivas took the stage with Intel CEO Lip-Bu Tan for a live demo — actually processing a confidential deal document on stage. What they showed off was the "hybrid local-server inference orchestrator." In plain terms: software that decides, in real time and mid-task, which AI workloads stay on your device and which get sent to frontier cloud models.
Here's how it works. A small local model running on your device acts as the router. It looks at each incoming task and evaluates whether it contains sensitive information, how much compute it needs, and whether it can be handled entirely on-device. Sensitive data — financial records, health info, personal files — stays local, while only the parts that need heavy reasoning get encrypted and sent to the cloud. Most real tasks are a mix of both, so the system splits a single job, processes both halves in parallel, and reconciles the results.
What makes the Windows version especially useful is Office integration. It plugs directly into Word, Excel, PowerPoint, and Outlook to read, write, and edit local files. On the cloud side, Claude Opus runs the orchestration across 19 models — complex reasoning goes to Claude Opus, web search to Gemini, coding to the GPT family, and speed-sensitive tasks to Grok. It's been confirmed running on hardware like Intel Core Ultra Series 3 and Nvidia RTX Spark, and the framework itself isn't tied to any particular model or chip.
But Should You Actually Trust That Promise?
CEO Srinivas laid out what winning this game looks like in an interview with CNBC: "Whoever balances accuracy, latency, cost, privacy, and intelligence together to maximize 'token value per watt per user' ultimately wins." Sounds great. But flip that sentence around and a different picture shows up.
Decrypt's take nails it. This strategy of not centralizing everything on servers is, first and foremost, an economic move to cut per-token costs for the company — the privacy upside is more of a side effect that happens to come along with it. And even if you want full offline isolation, you can't get it. Whatever traffic goes to the cloud always routes through Perplexity's own servers. That means you don't get true data ownership or full control.
The bigger issue is that this entire promise hinges on how accurately the local router classifies tasks. Winbuzzer points out that if task classification gets it wrong, sensitive content could leak off-device — or the router could over-classify things as sensitive and tank performance. And there's a relevant trust signal worth noting here: as of May 31, 2026, nine organizations have filed copyright and trademark infringement lawsuits against Perplexity. That track record is hard to ignore when you're trying to gauge how well a company will actually keep its data-handling promises.
| Traditional AI Assistant (Fully Cloud) | Perplexity Computer (Hybrid) | |
|---|---|---|
| Sensitive data handling | Always sent to servers | Local router decides, then processes on-device |
| Internet outage | Task stops | Local model can keep going |
| Full offline isolation | Not possible to begin with | Not possible — cloud traffic always routes through Perplexity's servers |
| Trust prerequisite | Just trust the vendor | Need both router classification accuracy AND vendor trust |
If your company is considering a team-wide rollout, there's one more layer to check. European outlet heise points out that businesses should review a Data Processing Agreement (DPA) and Standard Contractual Clauses (SCC) before adopting the tool. Unlike Microsoft Copilot, which stays within a single cloud ecosystem, this connects to 400+ external services, which makes the data flow itself far more complex.
This isn't a new kind of worry
Apple, Microsoft, Nvidia, and Google are all pushing in the same direction — on-device and hybrid inference. This isn't a Perplexity-only risk; it's more accurate to say the entire category of desktop AI agents is going through this same trust-verification phase right now.
If You're Going to Turn This On, Do It in This Order
- Check your subscription tier
Personal Computer opened up to Pro tier ($20/month) subscribers starting May 7. For the exact tier rollout timing on Windows, your account dashboard is the most reliable source. - Decide which folders never get connected
Keep original contracts, unreleased financial documents, and folders with customer personal data off-limits entirely until you've fully verified the router's classification accuracy. - Test with low-risk tasks first
Start with things like email triage or cleaning up meeting notes — stuff that wouldn't hurt much if it leaked — and watch a few runs to see what stays local versus what goes to the cloud. - Turn on Office integration
Once you're comfortable, connect Word, Excel, PowerPoint, and Outlook and start using it for real work — drafting from local files, analysis, file management. - Get the paperwork sorted before team rollout
If multiple people are going to use this, loop in IT and legal to review the Data Processing Agreement and Standard Contractual Clauses first.
Deep Dive Resources
Perplexity Brings Orchestrator Agent to Windows The original piece covering the product evolution timeline and Computex demo details. opentools.ai
Hybrid Local-Server Inference Orchestrator, Explained A technical breakdown of how the local router classifies tasks and what hardware it needs. marktechpost.com
Perplexity Computer Adds Local-Cloud Task Splitting Covers the Windows rollout timing and feature set. 9to5mac.com
Why Perplexity Wants Your Laptop to Do Some of the AI Work A critical look at the hybrid strategy through a cost-optimization lens. decrypt.co
Testing AI PC Privacy With a Local-Cloud Router Covers the lawsuit history and comparisons with competitors (Copilot+ PC, RTX Spark). winbuzzer.com
Perplexity Personal Computer Architecture, Explained Breaks down the 4-layer architecture for running a Mac mini as a 24/7 agent, plus practical tips. singhajit.com
