Perplexity's Portable Computer lands on Windows RTX

Perplexity puts its agent on a Windows desk

Perplexity has brought Portable Computer — its local-first agent — to Windows PCs running NVIDIA RTX graphics cards. The product first launched on Linux in late August; Windows support is now generally available on compatible NVIDIA GeForce RTX and RTX PRO workstations, according to NVIDIA’s 14 September announcement.

The framing from Perplexity’s side is blunt. VentureBeat quoted Nate, the company’s VP of engineering for infrastructure and enterprise, on what the launch was meant to deliver: We’ve basically brought the exact same UI to a fully local app.

The pitch is the one Perplexity has been making since August: run the agent on the box in front of you, keep the files it reads on that box, and ask the cloud for help only when the task genuinely needs it. NVIDIA spells it plainly — sensitive information stays on device and locally completed work doesn’t consume Perplexity Computer credits. The cloud bill lands only on the step that touches the cloud.

What’s local, what isn’t

The local side is long. The reasoning, planning, scheduling and local file indexing all run on the box and never round-trip to Perplexity’s servers, per the Layer3Labs breakdown.

The leaving side is short:

• Web searches go out as normal.
• Connector calls reach Google Drive, Gmail, Slack and GitHub on your behalf.
• An opt-in cloud reasoner, with your permission, takes a hard question and returns written guidance — with no access to local files or local tools.

That cloud escalation shares the question, not the data behind it. For a UK regulated buyer, that is the part a compliance review will spend the most time on.

The hardware floor just became realistic

When Perplexity first showed Portable Computer in August, the demo sat on NVIDIA’s DGX Spark — a desktop supercomputer priced around $4,700. That risked making the product feel like a research project. By early September, hands-on reviews confirmed a second supported path actually works: a Windows PC with a recent high-end NVIDIA card runs the same agent. The exact spec sits in the box below, but the takeaway is that the floor has moved from a niche workstation to a card many readers already own.

The catch is that the floor is exactly that. Running an agent plus its working memory plus tool-call context eats most of the card, and a 16GB consumer card is still locked out. A bare-minimum build leaves little headroom for the larger local models Perplexity is likely to ship next.

What to watch

The release matters less as a single product than as a template. For two years the local-or-cloud question has been a binary — trust a hosted agent with everything, or run a smaller model and accept that hard reasoning will fail. Perplexity and NVIDIA have now shipped a credible third path: a competent local model by default, with permission-gated cloud escalation for the steps that need it.

13.4ptsthe gap between a fully-local Portable Computer run and the same task with the cloud reasoner switched on, on Terminal Bench 2.1 — Perplexity’s own figures.

That gap is the clearest published measure of what local costs in reasoning. It is also the case for leaving the cloud reasoner on by default, and the next version of this product will live or die on closing it.

NVIDIA gets a local-AI platform story for RTX that is more concrete than another open-weight benchmark — the agent loop, the connectors, the credit economics, all running on its silicon (the duopoly thesis looks sharper for it). The interesting counterweight is Apple silicon, which Perplexity has said is not on Portable Computer’s roadmap. The company has a separate inference-engine project and Mac-side hybrid-compute work targeting the same end through different silicon, but a UK team with a fleet of Mac Studios is blocked from this particular build of the idea.

Three near-term signals worth watching:

• The local-model gap. The 13-point drop to the cloud reasoner is the headline problem to close. Expect the next model drop to be judged on whether it shrinks.
• The connector list. Every new connector is a place a UK regulated team’s compliance officer will ask where the query actually goes — and the answer is out to the third-party service, even though the reasoning stays on the box.
• The subscription price. Pro and Max are the entry tiers. A move on either is the most direct signal of how serious Perplexity is about the local-first story.

The broader signal is simpler than any of those. A credible vendor has shipped a local agent that escalates to the cloud only with permission, on hardware many readers already own or could afford. The category exists now. Watch how the rest of the field responds.