Perplexity has brought its AI agent, Portable Computer, to Windows machines running NVIDIA's RTX graphics cards. Until now, this local version only worked on Linux computers and NVIDIA's specialized DGX Spark boxes. Windows is where most office workers actually sit, so this is the version that will decide whether the idea catches on beyond hobbyists.
The pitch is simple. Instead of sending your documents to a company server to be read by an AI, the AI model runs on the graphics card already inside your computer. Perplexity uses an open model called Qwen, tuned specifically to work with its agent software. The agent can connect to Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub, and carry out multi-step tasks like sorting through pull requests, tracing tax drag across two years of brokerage statements, or digging through a spreadsheet to find where customers drop off.
For businesses that handle sensitive information, this solves a real problem. Law firms, accountants, financial advisors and healthcare providers have all been stuck choosing between using AI and following data rules that forbid sending client files to outside servers. A tool that reads a tax return on the machine itself, without that file ever touching a chatbot, removes that trade-off for a meaningful slice of daily work.
There is a real cost hiding in the pitch, though. This only runs on graphics cards with 24 gigabytes of memory or more, which means an NVIDIA RTX 4090, an RTX 5090, or one of NVIDIA's professional workstation cards. Those cards start around sixteen hundred dollars and run into the thousands for the workstation versions. A typical office laptop does not have anywhere near that much graphics memory. So this is not free for most companies; it is a hardware purchase disguised as a software update.
That hardware requirement also explains why NVIDIA is the one publishing this announcement, not Perplexity alone. NVIDIA benefits every time a software company builds a reason for businesses to buy a more expensive graphics card. The "local AI" push happening right now across NVIDIA's blog, from this Perplexity release to open models like DeepSeek and Qwen being tuned to run on NVIDIA hardware, is as much a sales strategy for NVIDIA as it is a privacy feature for users.
The bigger pattern matters more than this one release. OpenAI has its own AI-powered browser, Atlas, competing directly with Perplexity's Comet. Microsoft used its Build conference this year to shift its Copilot assistant toward running more work locally on the device instead of in the cloud. Every major AI company is now building a version of the same idea: keep the routine, sensitive work on your own machine, and only send the hard problems to the cloud when needed.
For a business owner, the decision is not whether to adopt this exact product. It is whether your team's hardware already qualifies, and whether the tasks you want automated, like reviewing contracts or reconciling accounts, are sensitive enough to justify the investment. If your designers or engineers already have powerful graphics cards for other work, this comes close to free. If not, you are buying new computers to get it, and that math needs to be done before anyone gets excited about the privacy pitch.