The Credit Counter That Never Moves

Perplexity's new AI agent runs entirely on a box on your desk. No tokens, no cloud bill, and your files never leave the room.

Welcome Back,

Perplexity launched Portable Computer on August 25, a version of its agentic "Computer" platform that runs entirely on hardware users already own, starting with Nvidia's DGX Spark desktop supercomputer and Linux machines equipped with Nvidia RTX GPUs, according to VentureBeat. The model, the user's files, and the work itself can all stay on the machine.

During Monday's press briefing, the most telling detail wasn't a benchmark score, it was a credit counter in the corner of the screen, sitting motionless at zero while the agent churned through a folder of tax documents. That single image captures the whole pitch better than any spec sheet, an AI agent doing real, complex work while costing nothing per task, because the computation never leaves the room.

Today we look at exactly what runs locally versus what still needs the cloud, the specific models making this possible, what the hardware actually costs and who it's realistically for right now, and why Perplexity and Nvidia both frame this as the start of a much bigger shift, not a niche side project.

🖥️ What Actually Stays On the Machine

Portable Computer is a local-first agent, according to Perplexity's own announcement. The entire stack runs locally by default, the model, harness, conversation, and trajectory all live on the user's machine. The orchestrator, planner, tool router, scheduler, and even the search index run directly on the user's own device, a genuinely complete inversion of how every previous Perplexity agent has worked, where requests were sent out, a cloud model performed the calculations, and the response came back.

Work that needs the outside world, web search, third-party connectors, or escalation to a stronger model in the cloud, is invoked only when necessary, and always gated by the user first. The company states plainly that sensitive data therefore never leaves the device without explicit permission, and local models carry no inference fee, the system is private and cost-effective by construction, not as an afterthought bolted onto an existing cloud product.

That "same UI" detail is genuinely important for adoption. Users don't have to learn a different interface or a stripped-down local-only mode, they get the identical agent experience, just running on a box they own instead of a server farm they're renting time on.

The orchestrator, planner, and search index all run locally, not just the underlying language model itself.

🤖 Small Models, Doing Genuinely Complex Work

Making a full agentic system run on a desktop box required genuinely efficient models, not scaled-down versions of frontier ones. Portable Computer uses small, efficient models such as Qwen 3.8 and NVIDIA Nemotron 3.5 Lightning, according to Android Authority, including PPLX 27B, a post-trained version of the Qwen 3.8 model built specifically for this purpose.

Dictation mode is a nice illustration of how far the local-only design goes, it runs entirely on-device using the NVIDIA Nemotron 3.5 ASR model, letting a user brief the agent while reading a document or sketching notes on paper, transcription and file actions all staying on the machine, without the audio ever touching the cloud, according to Perplexity's own blog post.

💰 The Real Cost, and Who This Is Actually For

It's worth being direct about the hardware barrier here before getting swept up in the pitch. An Nvidia DGX Spark costs somewhere between $4,000 and $4,700, a genuinely steep one-time purchase compared to, say, a Mac Mini at roughly $900, which has been the go-to budget option for running local AI agents until now.

Access itself is also gated, currently limited to paid Perplexity accounts, Pro, Max, Enterprise Pro, and Enterprise Max subscribers, and the software currently runs only on Linux, with Windows support promised for September. The honest read here, echoed across multiple outlets covering the launch, is that Portable Computer remains, for now, an offering aimed at a genuinely niche audience, people who already own compatible hardware or are willing to make a serious one-time investment to get there.

A compatible RTX GPU with 24GB or more of VRAM is the realistic entry point for most people beyond the dedicated DGX Spark.

🧩 Why Not Just Run Qwen in Ollama Yourself?

This is a genuinely fair question, and Perplexity has a specific answer for it. According to Perplexity, Portable Computer "packages step-level routing with the agent harness, local models, inference, tools, app connectors, and sandboxed execution in one maintained system," as How-To Geek's coverage puts it.

That's the real value proposition worth understanding, the underlying open model, Qwen 3.8, isn't the hard part anymore, plenty of people already run models like it locally through free tools. What Perplexity is actually selling is the surrounding infrastructure, connectors to Google Drive, Gmail, Slack, and GitHub, sandboxed execution for safety, and a maintained, integrated system that doesn't require the user to stitch together their own agent harness from scratch.

Security is designed to match the cloud version of Perplexity Computer directly, code and tool execution run in isolated sandbox environments with controlled access to files and connected apps, a genuinely important detail for anyone trusting an autonomous agent with access to real, sensitive files and services.

📈 A Bet on Where the Chips Are Heading

Both companies frame this launch as something bigger than a single niche product. Perplexity's researchers describe it as part of "a broader shift in which increasingly capable agents move from remote infrastructure to individual and local devices," with both companies betting that advances in chips and open models will keep expanding what a box on a desk can realistically do.

Perplexity's own reasoning is direct about the compounding advantage they expect from this bet, "As models get stronger and chips get faster, more people will run complex workflows on their own machines," the company wrote, framing today's hardware barrier as a temporary state rather than a permanent limitation.

Perplexity and Nvidia are betting that today's hardware barrier shrinks as chips and open models keep improving.

🧠 AI Spotlight Analysis

For two years, the AI industry has measured its ambitions in gigawatts and tokens per dollar, ever bigger data centers, ever more compute purchased at scale. Portable Computer proposes a genuinely different meter, one that never runs, at least for the portion of work handled locally. That's a real philosophical departure from where most of the industry's headlines have been pointed.

It's also a notable strategic move for Perplexity specifically, a company one outlet bluntly described as "a largely forgotten player in the AI space save for its role as Joe Rogan's personal fact checker." Partnering directly with Nvidia on hardware-level integration, rather than just building another chatbot wrapper, is a genuine bid for relevance in a space where product differentiation has become genuinely difficult.

That image is genuinely more persuasive than a benchmark chart would be, because it speaks directly to the two things people actually worry about with AI agents, cost and privacy, rather than raw capability. Whether Portable Computer becomes a meaningful product depends less on today's $4,000 hardware requirement and more on whether Perplexity's bet about falling chip costs and improving small models plays out the way the company expects.

🔗 Sources and Further Reading

Thanks for reading AI Spotlight.

Our mission is simple: deliver clear, trustworthy, and actionable AI insights that help professionals stay ahead without the hype.