Perplexity Just Made Your AI Agent Free to Run. There's a Catch.

I ran a quick calculation last week. One of my consulting clients was spending $1,200 a month on AI agent calls. Document reviews, data pulls, report drafts. All hitting cloud APIs, all billing by the token.

That number is about to look optional.

Perplexity just teamed up with Nvidia to launch something called Portable Computer. It's an AI agent that runs entirely on your own hardware. No cloud. No token costs. Zero.

How It Actually Works

Portable Computer packages a local AI model, an agent system, app connectors, and a security sandbox into one application. It runs on Nvidia GPUs. You need at least an RTX 3090 with 24GB of video memory. At launch, it ships with two models: Qwen 3.8 27B and Perplexity's own PPLX 27B.

The agent handles real work. It reviews financial documents, analyzes CSV data, connects to Google Drive and Gmail, and pushes results to Slack. Think of it as a capable junior analyst sitting on your desk. One that never sends your data anywhere.

The Privacy Angle Matters

This part caught my attention. Portable Computer runs a PII classifier that screens anything before it leaves your machine. It defaults to local execution and only reaches out to a cloud model (like Claude Opus 5) if you explicitly give permission.

For businesses in healthcare, legal, or finance, this solves a real problem. You get AI agent capability without sending client data through someone else's servers. That compliance conversation with your legal team just got shorter.

So What's the Catch?

A few things.

First, you need hardware. An RTX 3090 or newer with 24GB of VRAM isn't sitting in every office. The sweet spot is Nvidia's DGX Spark desktop, but that's a real investment. You're trading recurring cloud costs for upfront hardware costs.

Second, performance is good but not great. On Perplexity's internal benchmarks, the local models scored 82-85% on knowledge work tasks. For coding tasks, local performance hit 59.6%. Add cloud escalation and it jumps to 73%, but then you're paying again ($0.415 per task on average).

Third, it only runs on Linux right now. Windows support is coming in September. If your team runs Windows, you're waiting.

And you need a Perplexity Pro, Max, or Enterprise subscription. This isn't free software on free models. It's a premium feature for paying customers.

Why This Matters for Your Business

AI agents consume way more tokens than a simple chatbot conversation. Every time an agent takes a step, reasons about a task, reads a file, or writes output, that's tokens burning. A single complex task can eat through thousands of tokens across dozens of steps.

Running those steps locally changes the math completely. For businesses that run repetitive agent workflows like document processing, data analysis, and regular reporting, the savings add up fast.

Take my client spending $1,200 a month. Most of those tasks are exactly the kind Portable Computer handles. Document reviews, data pulls, routine analysis. If even half of those ran locally, that's $600 a month back in the budget. Over a year, that's $7,200, which buys actual hardware.

What I'd Do Right Now

Don't rush out and buy a GPU. But do this: look at your AI spending and figure out which tasks are repetitive and don't need the absolute best model. Those are your candidates for local execution.

Cloud AI will always exist for the hardest problems. But for the daily grind of business agent work, local execution is becoming a real option. Perplexity just made it a product instead of a hobbyist project.

That shift from hobby to product is the real story here.