Analysis · 2026-09-01 · 9 min read

Perplexity and NVIDIA Launch Portable Computer: Zero-Token Local AI

Perplexity and NVIDIA shipped Portable Computer, an agent platform that runs entirely on your own GPU — no per-token billing, no metered API. Here's what it costs in hardware and electricity, and when local beats the cloud.

TL;DR

Perplexity and NVIDIA launched Portable Computer, a local-first version of Perplexity's agentic Computer platform that runs entirely on your own hardware — no per-token charges, no metered billing.
It targets NVIDIA DGX Spark and high-end RTX cards, and needs a GPU with 24 GB or more of VRAM. Linux is available now; Windows is coming.
It runs open-weight models — Alibaba's Qwen family among them — instead of proprietary hosted endpoints.
Cloud fallback is opt-in: the app asks permission before shipping a heavy reasoning task off-device, so your files and documents stay local by default.
The real economics: you swap a variable per-million-token bill for a fixed capital cost plus electricity. Break-even depends almost entirely on your monthly token volume.

!Perplexity

!NVIDIA

!Qwen

1. What Portable Computer Actually Is

Perplexity's Computer platform is an agentic runtime: it browses, reads files, writes code, and chains tool calls until a task is finished. Every one of those steps normally costs tokens on a hosted frontier model, and agentic loops are the single most token-hungry workload in production AI.

Portable Computer moves that loop onto your machine. The agent, the model weights, and your data all sit behind the same firewall — your own. NVIDIA supplies the hardware target and the inference stack; Perplexity supplies the agent orchestration layer that people already know from the cloud product.

The pitch is short: the marginal cost of a task goes to zero.

2. How Zero-Token Local AI Works

The escalation prompt is the underrated design decision. Most "hybrid" local setups silently route hard prompts to the cloud, which quietly recreates the bill and the data-exposure surface you were trying to escape. Making it a consented, visible action turns cloud usage into a budget you can actually watch.

3. Hardware and System Requirements

Workstation: built for NVIDIA's DGX Spark or a powerful RTX graphics card.
Memory: a GPU with 24 GB VRAM minimum. Below that, useful agentic context windows don't fit alongside the weights.
Operating system: Linux available now, Windows support arriving soon.

24 GB is the practical dividing line. It's enough for a quantized mid-size open-weight model plus a working context, which is exactly the shape of an agent loop: modest model, long scratchpad.

4. The Real Cost Math

"Zero token cost" is true at the margin and false at the ledger. You're trading a variable cost for a fixed one. Here's the comparison that matters.

Cloud side. A typical agentic workflow burns far more input than output — tool results, file contents, and prior turns get re-read every step. Assume 20M input and 2M output tokens per month for one heavy user:

Local side. A 24 GB-class RTX workstation card plus the rest of the machine lands in the $2,000–$4,000 range; a DGX Spark-class box sits well above it. Electricity for a 350 W card running four hours a day at $0.15/kWh is roughly $6–$7 a month.

So the break-even against a frontier flagship at that volume is roughly 12–20 months on hardware alone — and it collapses to well under a year if you have multiple users on one box, or if your agent loops are heavier than the example. Against cheap open-weight hosted endpoints, the hardware may never pay for itself on cost alone. Local wins on privacy and predictability long before it wins on price.

Run your own numbers against live rates in the cost calculator, model an agent loop in the Agent Loop Cost Estimator, and compare hosted alternatives in the pricing table.

5. Who Should Actually Run This

Strong fit

Teams whose source code or client documents legally cannot transit a third-party API.
Developers running long, repetitive agent loops where the same context is re-read hundreds of times a day.
Anyone who needs a fixed, forecastable AI line item instead of a bill that scales with enthusiasm.
Offline and air-gapped environments.

Weak fit

Low-volume users. A few hundred thousand tokens a month costs less than lunch on a hosted open-weight model.
Workloads that genuinely need frontier reasoning on every call — you'll hit the cloud fallback constantly and pay anyway.
Teams without a 24 GB+ GPU and no appetite to buy one.

6. What This Signals for Token Pricing

The broader story is that the token is becoming an optional unit of account. Hosted per-token pricing assumes inference has to happen in someone else's data center. Open-weight models good enough for agentic work, running on hardware that already sits under a desk, break that assumption for a growing slice of workloads.

That doesn't kill hosted APIs — frontier reasoning still lives in the cloud, and will for the foreseeable future. What it does is cap what the routine tier can charge. When the boring 80% of agent steps can run for the price of electricity, providers can't price those steps like frontier reasoning. Expect continued pressure on the cheap end of every provider's lineup.

Bottom Line

Portable Computer is the clearest consumer-facing bet yet that a lot of agentic work doesn't need a data center. If your constraint is privacy, predictability, or a genuinely heavy agent loop, a 24 GB GPU turns a scary variable bill into a known fixed one. If you're a light user, the hosted open-weight tier is still cheaper than the electricity plus the hardware. Either way, the interesting number is no longer just dollars per million tokens — it's dollars per finished task, and local hardware just entered that comparison seriously.