Perplexity Wants to Run Your AI Work on Your Own Hardware

Some of the AI work people actually trust is also the work people don't want to send to someone else's server: proprietary code, draft term sheets, internal documents. A new wave of tools wants to let you run that work on your own machine, with the cloud as an opt-in escalation only when a task needs it. Perplexity's Portable Computer, built with NVIDIA, is the most visible bet on that local-first category so far. The question is whether the trade-off actually holds: data control and no per-credit charges on local work, in exchange for capable hardware and a model ceiling below frontier cloud.

"Local-first AI" means the model runs on the user's own hardware rather than on a vendor's servers; the cloud only steps in when a task needs more than the local box can deliver. Perplexity's Portable Computer is a stripped-down, fully local version of the company's hosted "Computer" agent: the orchestrator, planner, tool router, scheduler, durable task queue, and a local search index all live on the device. The default model is either Qwen 3.8 27B or PPLX 27B, a post-trained variant of the Qwen model, with NVIDIA's Nemotron 3.5 Lightning, a 30B open-weight model, coming to the model picker. In practice, a user can hand the agent a private repository, a draft merger agreement, or an internal incident postmortem and have the tool review, summarize, and route tasks across files without any of that text leaving the desk.

The hardware anchor is the catch. Portable Computer runs on NVIDIA's DGX Spark, a desktop-class AI workstation, and is set to expand to NVIDIA RTX GPU PCs. NVIDIA's own framing of the product is an agent "optimized for DGX Spark," which reads less like a cross-platform consumer product and more like a single-vendor showcase on a specific box. On-device work does not burn through Perplexity's subscription credits, and cloud calls only happen when the user explicitly authorizes them: a meaningfully different default from a hosted agent that phones home for every step.

That is the upside. The cost of that split is harder to see. The local model is a 27B-class open model, capable but not at the frontier; tasks that genuinely need a top-tier model still escalate to the cloud, and the user has to know enough to trigger that escalation. The hardware is also real money: DGX Spark sits in the multi-thousand-dollar desktop-workstation tier, and that puts a hard ceiling on who can actually run the "own the stack" pitch that Perplexity's product page sells. It is also a real test for the credit-based AI business model. If more high-volume, low-stakes work moves onto local boxes, vendors have to make the cloud calls they do sell both rarer and more obviously worth paying for.

Skepticism is already surfacing in the Hacker News thread on the announcement. The recurring questions are about who the target user is, given that DGX Spark buyers are by definition technical; whether the bet makes sense as a category move or just a hardware-anchored demo; and how the Pro/Max subscription framing squares with software that is supposed to escape the cloud. Those are not fatal objections, but they are the falsifier the launch has not yet passed: a single-vendor product on a niche workstation does not yet prove that local-first AI is a mass-market shift.

What Perplexity is actually doing is unbundling the cloud-AI default into two parts the user can mix: a local agent that handles the high-volume, private-by-default work, and a cloud escalation that only fires when the local model hits its ceiling. The watch items over the next quarter are the RTX GPU PC rollout, which is the first test of whether this category can run on hardware that is not a multi-thousand-dollar workstation, and whether any of the open-weight models in the picker get independently benchmarked against hosted frontier models on the kinds of tasks Perplexity is advertising.