# Perplexity brings local AI to Windows—but your graphics card needs 24GB of memory

> Portable Computer can handle work on a compatible Windows PC and ask before using cloud AI. Here is what the hardware requirement and privacy promise mean.

- Canonical URL: https://techxtelco.com/news/perplexity-local-ai-windows-rtx/
- Published: 2026-09-15
- Updated: 2026-09-15
- Topics: AI, News
- Author: Michel Elijah
- Image: https://anrkjmgqmwfosbvyzrzc.supabase.co/storage/v1/object/public/article-images/perplexity-local-ai-windows-rtx/hero-v1.webp
- Publisher: Tech X Telco (https://techxtelco.com)

Perplexity’s Portable Computer is coming to a more familiar place: a Windows PC. NVIDIA announced Windows support on 14 September, extending the local AI agent beyond its earlier Linux and DGX Spark release. The catch is substantial: it requires a compatible GeForce RTX or RTX PRO graphics card with at least 24GB of video memory.

### What you need to know

Check your graphics card before planning around this release. A PC advertised with 32GB of RAM does not necessarily have 24GB of graphics memory. They are different specifications.

## What the Windows release adds

According to NVIDIA, the agent can work across files and complete multistep tasks using a local model. Work completed locally does not consume Perplexity Computer credits. More demanding work can use cloud models, with permission requested before information leaves the device. NVIDIA also describes app connectors and a sandbox for tools.

The practical appeal is less copying and pasting between a folder of documents and a chatbot. Whether that saves time depends on the job: assembling a reliable answer from several files is useful; producing a polished answer that still needs every figure reconstructed is less so.

## The 24GB detail to check before spending money

VRAM is memory associated with the graphics processor. Ordinary system RAM is a separate resource. Adding more system memory does not, by itself, satisfy a requirement for a graphics card with more VRAM.

Look up the exact graphics-card model in your machine and its dedicated memory specification. A retailer’s broad “AI PC” label is not enough evidence of compatibility. Neither is knowing that the computer has an NVIDIA card: the model and memory capacity matter.

Our advice is to confirm the app’s current hardware and account requirements before buying anything. This announcement is not a reason on its own to replace a perfectly serviceable laptop.

## Local processing is a useful boundary, not a universal guarantee

For a first trial, use copies of a few non-sensitive documents in a separate folder. Ask for a summary that identifies the filename and page supporting each important claim. Keep the originals untouched and compare the output with them. This is a suggested evaluation, not a workflow we have tested in Portable Computer.

Treat permission to read a folder, permission to send information to a cloud model, and permission to change something in a connected service as separate decisions. If a proposed action is unclear, stop and inspect it before agreeing. A local model does not make a mistaken email, overwritten file or incorrect calculation harmless.

## What remains to check

We could read NVIDIA’s launch announcement, but Perplexity’s linked Windows announcement blocked access during verification. We have therefore left subscription eligibility, Australian pricing and the exact model offered in the Windows app unconfirmed. We have not measured speed, accuracy or network traffic.

For someone who already owns suitable hardware, this is worth assessing with a small, repeatable task. The useful test is whether it produces work you can check and keeps you in control of the next action—not simply whether the model runs on your desk.

### Sources and further reading

- [NVIDIA: Portable Computer comes to Windows](https://blogs.nvidia.com/blog/local-ai-perplexity-windows-pcs/): Windows launch; compatible RTX GPUs with 24GB or more VRAM; permission before cloud processing; locally completed work does not consume Computer credits