Perplexity Brings Its AI Agent to Windows PCs With NVIDIA RTX GPUs

What happened

Perplexity has made Portable Computer available in its Windows app on GeForce RTX PCs and RTX PRO workstations, and the competitive shift it signals is unmistakable: the enterprise AI agent runs on the PC first and in the cloud second. The announcement comes from NVIDIA's blog^[1], which presents the product as the local version of the Perplexity Computer agent. The release extends support already live on DGX Spark systems and on RTX PCs running Linux.

The hardware requirement is specific: an NVIDIA GeForce RTX or RTX PRO GPU with at least 24 GB of video memory. Support for DGX Station is listed as coming soon, so it should be read as a promise: the product shipping today is the Windows one.

This is the clearest signal yet of a change in how enterprise AI gets distributed.

The default moves from the cloud to the employee's machine, with the cloud demoted to an escalation tier. The software vendor sells an agent, and the chip supplier collects on the prerequisite.

What the product really is

Behind the product language sits an agent that plans and executes multi-step tasks on the PC.

It uses local models to analyse data, stitch together information scattered across files and handle recurring work. Everything runs on the machine's GPU, with sensitive data staying on the local disk.

The package includes an integrated browser and a proprietary sandbox called SPACE. The default local model is Qwen 3.8 27B, post-trained for Perplexity Computer and optimised for RTX GPUs, as NVIDIA's post^[1] notes. The user is spared both model selection and software stack configuration.

The most consequential commercial detail fits in a single line: work completed locally consumes zero Perplexity Computer credits. When a task demands more reasoning capacity, the agent proposes escalating to the cloud and asks permission before any data leaves the machine. Connectors to Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub hook the agent into workflows already in use.

From cloud-first to on-device-first

The axis of competition is shifting.

For two years, vendors have sold AI capability as a metered service: tokens, credits, enterprise contracts with data residency written in as a clause. Perplexity reverses the order and puts the local machine first.

The market signal is this: value is migrating from the model to the deployment layer. The model is a commodity, as the drop in frontier pricing with every release shows. What remains defensible is the point where the agent meets the company's data, and that point now sits on the desktop.

The examples NVIDIA cites speak to three business functions. Engineering: reviewing open pull requests on GitHub, sorted by status, with proposed documentation updates. Finance: analysing two years of statements and tax filings, with every figure cited to file and page, and with the documents never leaving the PC.

The third case, a startup analysing its funnel export locally and publishing the results to Slack, marks out the perimeter: proprietary data, repetitive work, output delivered in the channels the team already uses.

The GPU as a lock-in gate

This is where the nature of the graphics card changes.

Until yesterday, the GPU was a cost line in the IT budget, chosen by procurement on price and availability. From today it becomes the commercial prerequisite for a software product: whoever wants the local agent buys NVIDIA.

The 24 GB video memory threshold excludes most of the installed corporate fleet. That turns the launch into a hardware refresh programme dressed up as a software release.

The lock-in has two layers.

The first is the model post-trained and optimised for RTX, which makes porting the same agent to different silicon expensive. The second is operational habit: once finance and engineering work with an agent that sees their files, switching vendors means redoing connectors, permissions and training. The supplier with the deepest deployment captures more durable revenue than the supplier with the highest benchmark.

Who feels the impact

The direct implications reach Microsoft, OpenAI and Anthropic on the agent front, and AMD, Intel and Apple on the silicon front. Microsoft sells Copilot inside Windows, and now sees a third-party agent using the Outlook, OneDrive and Word connectors on the same machine. The battle for the enterprise desktop opens up inside Redmond's own operating system.

For OpenAI and Anthropic, the release raises a roadmap question. Their agents live in the cloud and monetise by consumption: a competitor that zeroes out credits for local work puts pressure on pricing for routine tasks.

For alternative chipmakers, the message is blunt.

An enterprise agent optimised for RTX reinforces NVIDIA's moat outside the data centre too, on every analyst's PC. AMD, Intel and Apple silicon Macs are shut out of this product, and NVIDIA's post says nothing about plans to open up to other platforms.

The governance trade-off for the CDO

For the Chief Digital Officer evaluating AI agent vendors, the move redefines the compromise. On the benefits side sits governance: sensitive data stays on the machine, escalation to the cloud requires explicit consent, and metered costs fall for recurring tasks. That is a strong argument with legal and with internal audit.

On the cost side sits the technology constraint.

The NVIDIA prerequisite ties the agentic strategy to one hardware supplier and to a very specific tier of machines. Data sovereignty is a product, and here the product ships with a graphics card included.

The three-year question is simple: how much is it worth to the board to know for certain that a tax document stays on the PC? When that value exceeds the price of the constraint, the product enters the portfolio. It falls to the CDO to put both weights on the same scale before the next licence renewal.

What to decide in the next 90 days

The Chief Strategy Officer needs to open a conversation with their agent vendor and ask for the on-device roadmap: who has a ready answer and who defers everything to the cloud. The most urgent partnership is with whoever controls the local deployment layer, because that is where the lifetime contract gets decided.

The CFO needs to revisit two line items. The first is the cloud credit budget for agents, which can come down for routine tasks. The second is the PC refresh plan, which risks going up because of the video memory threshold.

The Chief Digital Officer needs to launch a pilot in a function with sensitive data, finance or legal, and measure how many tasks stay local. The Technology Investor reads confirmation of a thesis: compute is gaining value while the model depreciates, and the GPU-equipped PC enters the perimeter of enterprise AI investment.

The market has moved.