Perplexity's Portable Computer runs AI agents locally on Windows - but you need a $2,000 GPU

Perplexity brought its Computer agent to Windows this week, running entirely on the machine instead of in the cloud. It launched on September 14 through the Microsoft Store, and it can review GitHub pull requests, dig through brokerage statements for tax-optimization opportunities, or watch a signup funnel and post the drop-off points to Slack, without any of that data leaving the PC. There's a catch that guts the announcement for most businesses: it only runs on an NVIDIA RTX or RTX PRO GPU with at least 24GB of VRAM.

What happened

According to NVIDIA's own announcement, Portable Computer is the local version of Perplexity's existing Computer agent, and it now ships inside the Perplexity Windows app. It runs Qwen 3.8 27B, a model NVIDIA says was "post-trained to work with Perplexity Computer and optimized for NVIDIA RTX GPUs," entirely on-device. Support for NVIDIA's DGX Station is coming next.

The feature is gated behind Perplexity's paid tiers — Pro ($20/month) or Max ($200/month) — and distributed through the Microsoft Store rather than a standalone installer.

What is genuinely new

Most agentic AI tools ship every document, email, and spreadsheet to a cloud model to reason over it. Portable Computer does the reasoning on the box itself, and Perplexity says completed local work doesn't draw down the cloud-based "Computer" credits that meter the rest of the product. For anything that needs heavier reasoning than the local model can manage, it asks permission before sending data off-device to a cloud model — an explicit consent step, not a silent fallback.

It also comes with real connectors out of the gate: Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub. NVIDIA's own examples are concrete rather than aspirational — reviewing and triaging pull requests by status, flagging documentation that's gone stale, and reading tax returns to spot fee or deduction opportunities.

What it means for a business owner

If your work involves data you don't want touching someone else's servers — client financials, contracts, anything under an NDA or a compliance regime — a local agent that can still read your inbox, your drive, and your GitHub repo is a real category, not a gimmick. The task list NVIDIA published (funnel analysis posted straight to Slack, PR triage, statement review) maps onto real ops and finance workflows, not demo filler.

The mechanism is also worth noting on its own: an agent that defaults to local execution and asks before escalating to the cloud is a more defensible privacy posture than "trust us with your data," and it's the kind of design choice worth asking any vendor pitching you an AI agent about.

The honest caveat

24GB of VRAM is not a spec most business hardware has. That tier starts around an RTX 4090 or 5090 — cards that run $1,600 to $2,000-plus on their own, before the rest of the machine. Most office laptops and desktops, including plenty bought in the last two years, don't have a discrete GPU at all, let alone one in that class. Tom's Hardware called the requirement out directly: this puts Portable Computer out of reach for the large majority of Windows users, full stop.

So today this isn't a business automation tool you deploy across a team. It's a feature for the one person in the building who already owns a workstation-class GPU, probably because they do 3D rendering, local model fine-tuning, or something similarly niche. Add the $20–$200 monthly Perplexity subscription on top, and the actual addressable audience for this specific launch is small.

What to do about it

Don't buy hardware to chase this. What's worth tracking is the design pattern, not this specific release: local-first execution with an explicit, visible handoff to the cloud only when needed. If you're evaluating any AI agent vendor for work that touches sensitive data, ask them directly where processing happens by default and what triggers a trip off-device. That question will matter more as GPU requirements come down — and they will, the same way every "needs a powerful GPU" AI feature has over the past few years.