共有:

Local-First AI

Perplexity Teams Up With NVIDIA
for an Agent With Zero Token Cost

Perplexity has unveiled Portable Computer, a fully local version of its agent platform, built with NVIDIA. Work handled entirely on-device carries no token charge at all, and the system asks permission before sending any step to a cloud frontier model. Here's what changed, and who it's actually for.

AI Navigate Editorial2026.08.267 min read

Task starts Runs on-device cost $0 Send to cloud? needs permission Frontier model
01

What Happened

Bringing "Computer"
Entirely Onto Your Machine

Perplexity announced, in its official blog post, a new version of its existing agent platform "Computer" — called Portable Computer — that runs entirely on hardware users already own. Per VentureBeat's reporting, the NVIDIA partnership means it starts with NVIDIA's DGX Spark desktop supercomputer and Linux machines with RTX GPUs. The local model, inference engine, tool sandbox, and app connectors ship as one packaged system that stays on-device, and every task begins locally by default.

The hardware bar is an NVIDIA GPU with at least 32GB of VRAM. At launch, users can choose Alibaba's Qwen 3.8 27B or PPLX 27B, a version Perplexity post-trained on its own agent harness, with NVIDIA's Nemotron 3.5 Lightning coming soon.

02

Portable Computer, by the numbers

32GB+
required VRAM (NVIDIA GPU)
27B
local model size (Qwen/PPLX)
$0
token cost for local work

Before it reaches the cloud,
it gets handled right where you are.


03

Why It Matters

A break from "the cloud handles everything"

Until now, Perplexity Computer orchestrated 20-plus models in the cloud, which meant per-use API charges were simply assumed.

Agent products up to now assumed that any capable reasoning meant running a model in the cloud, with token charges attached to every use. Portable Computer is a bet that a 27B-class model is now good enough to handle real agentic work on a local GPU. As MarkTechPost notes, it also isolates tool execution inside an OS-enforced sandbox with reduced privileges — and when work stays local, the cost of getting a result drops to effectively zero. It's one of the first concrete signs that the old assumption — send everything to a giant cloud model — is starting to crack.

04

Who It Affects

Who feels this, and how

Developers with RTX/DGX hardware

You can iterate freely without worrying about per-call API costs. But the 32GB+ VRAM bar rules out most consumer laptops, so check your actual hardware first.

Companies that can't send data off-premises

Keeping processing on-device fits well with enterprise or on-prem needs around sensitive data. Since cloud escalation requires explicit permission, the risk of data leaving accidentally is reduced.

Casual users on the cloud plan already

Without qualifying hardware, there's little upside here. Windows support isn't arriving until September, so for now this is limited to Linux users with the right NVIDIA GPU.

05

What to Do Next

If you're considering it

01

Check your GPU's VRAM

Confirm you actually have 32GB or more on an NVIDIA GPU (DGX Spark or an RTX card). Without it, there's no benefit to chase.

02

Confirm your subscription tier

Availability is currently limited to Pro, Max, Enterprise Pro, and Enterprise Max subscribers — the free tier appears to be excluded.

03

Decide your cloud-escalation policy up front

Any task that can't finish locally will prompt for permission to send it to a frontier model — set your approval criteria as an internal policy before rollout, not after.


06

Risks & Open Questions

Don't over-trust it yet

"Zero cost" is an appealing headline, but a few caveats apply. First, a 27B-class local model won't always match a top-tier cloud model's accuracy — complex reasoning or long-context tasks will still escalate to the cloud, where normal billing applies. Second, the 32GB+ VRAM requirement is not trivial, so the users who actually see "zero cost" benefits are largely developers and companies who already own qualifying GPUs. Third, Windows support doesn't land until September, so for now this remains a Linux-first offering with a narrower reach than the "local-first for everyone" framing might suggest.