Perplexity’s new local-first AI agent is designed to address two major concerns in the AI market: high token costs and privacy. At the same time, the product underscores a broader trend toward edge AI while also revealing a key obstacle — the high cost of the AI computer required to run it.

On Tuesday, the AI search vendor introduced Portable Computer, a local version of Perplexity Computer. Computer, first introduced in February, is a general-purpose digital worker that helps different AI models carry out autonomous workflows.

With Portable Computer, users run Computer on a personal device, particularly Nvidia DGX Spark, a compact desktop supercomputer. Perplexity built the local agent with Nvidia, and it runs on the personal AI computer using the Qwen 3.8 27B model from Chinese tech giant Alibaba, or PPLX 27, a post-trained version of Qwen.

Portable Computer reflects the growing movement toward agentic AI running on edge devices to reduce token costs and protect user privacy. Since the release of OpenClaw in 2025, vendors have increasingly focused on models or agents that can run on users’ devices. In June, Microsoft introduced its family of Aion models, designed to orchestrate local agentic workflows. In March, Anthropic updated Claude to access local computers, and open source AI lab Ai2 introduced a computer-use agent for desktop users.

That shift matters because more AI workloads are being evaluated not just on capability, but also on where data is processed and what it costs to keep agents running. Local execution can reduce the amount of data sent to external systems, but it also moves some of the burden to the device itself, which makes hardware and setup decisions part of the product experience.

“[Portable Computer] is another good example of the evolution toward local AI, enterprise AI being more important, private AI, people really caring about where does my AI run, what data can it access,” said Said Oouisaal, CEO and founder of edge computing vendor Zendeda. “You want to own your intelligence, and this is a good example of that.”

Perplexity said confidential data remains local and on-device, and that the system does not use many tokens or credits. Most tasks are handled locally on the computer, though some are escalated to the cloud when browser access, connected apps, or other frontier models are needed for advanced reasoning. That split highlights a practical limitation of local-first systems: even when the core workflow stays on-device, some enterprise tasks still depend on cloud capabilities.

Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget, said the approach may also help Perplexity reach new customers.

“It allows them to expand their customer reach a little bit,” Su said. “[They’re] using this as a gateway for people who are more cost-conscious or have a lot of sensitive documents that they are not willing to upload to the cloud.”

For Nvidia, which helped build the local agent, the product is also a way to expand the use of its DGX devices.

Some Problems

Although Portable Computer can help users be more conscious of token spending, it still comes with limitations. Users may save on tokens, but they face significant hardware costs with DGX Spark. A single DGX Spark computer costs $4,699.

“It’s not cheap,” Su said. “To buy DGX is not a random thing. It’s a very deliberate decision, and I do not think anyone is doing that lightly … even enterprises.”

If Perplexity wants to sell an approachable technology that enterprises can adopt, it likely needs to work with Microsoft and the Windows ecosystem, he said.

Another challenge is the complexity users face when working with an AI computer.

“It’s still quite complex to set up a system to do inference on your own,” Oouisaal said. He added that while it is easy to buy the hardware, it is harder to install the software layers needed, such as Linux or container software like Kubernetes, so locally created and run AI agents can function properly. He said that while Portable Computer is a single-function system, enterprise environments will require more expertise to serve different departments and agents.

“It’s not easy,” Oouisaal continued. “You need to be an engineer.”

Su said it remains unclear whether Portable Computer will reach the enterprise market in significant numbers in its current form.

“Enterprises are already really concerned with shadow AI and are already restricting machines that have internet access to all these [models],” he said. He added that although Perplexity and Nvidia can say only token-consumption data is transferred to their servers and all other data stays on-device, enterprises may still be wary of giving employees more independent access to AI.