Product Independently confirmed 2 publishers 3 min read Published Aug 25, 2026, 6:46 PM PDT

Perplexity makes "ask before the cloud" a feature, and prices it at DGX Spark

Portable Computer keeps agent work on hardware you own and asks permission before any single step escalates. The entry fee is 24GB of VRAM and a paid subscription.

The Product Desk

Send a correction

What happened

Perplexity announced Portable Computer on Tuesday, a version of its agentic Computer platform that runs on Nvidia DGX Spark and other Linux machines with compatible Nvidia GPUs.
Tasks start on the local machine, and the system asks permission before sending any individual step that needs stronger reasoning to a cloud frontier model.
The release is Linux-first and needs Nvidia hardware with at least 24GB of VRAM, including DGX Spark or DGX OS and Ubuntu machines on ARM or x64.
Model choice narrows to two at launch, Alibaba's Qwen 3.8 27B or Perplexity's post-trained PPLX 27B, against the 19 models Computer was pitched with.

Why it matters

What's now possible capability Procurement can now buy an answer to "where does our data go" as a device plus a consent step, instead of accepting a vendor promise about server-side retention.
What it costs cost The buyer pays roughly $4,500 up front, plus a paid Perplexity tier, before finding out whether a 27B local model handles enough of the work to justify it.
What doesn't add up contradiction Gizmodo describes Computer as an agent that works while you are away; the privacy control here needs someone present to approve each handoff, so the two pitches pull against each other.
Who's exposed exposure Gmail, Drive, Slack and GitHub connectors give data a route off the machine that no escalation prompt governs, which is where a security review will land first.

Next story · Product · 10h ago

Reviewers say Microsoft's $5,900 Surface Laptop Ultra suits buyers who will tinker with local AI

Confirmed3 publishers

The escalation prompt is the product. Everything beneath it, the inference engine, the agent harness, the sandbox, the app connectors, is assembly work a competent team could do itself and normally has to ^[8]^[7]. What a team cannot easily build is the moment where the agent halts, says this step exceeds the local model, and waits to be told yes or no ^[2]^[6]. That is a governance artifact dressed as a UX detail.

The billing design is the part worth reading twice. Locally completed work consumes no credits, and only the portion sent to the cloud counts toward cloud usage ^[10]. Run that backwards and the invoice becomes a meter: credits accrued are a rough measure of how much of a month's agent work left the box. For an auditor asking where the data went, a line item beats a policy statement.

The awkwardness is in the pairing. Computer was pitched in February as an always-on agent that keeps working whether or not you are at the keyboard, with the processing in the cloud ^[14]. Portable Computer's privacy claim depends on a human being present to approve handoffs ^[6]. Neither account says what happens to an unattended overnight run that hits a step the 27B local model cannot handle: it either stalls or the approval degrades into a reflex click. That distinction decides whether this is a compliance control or a dialog box.

The GPU is also not the perimeter. Portable Computer reads local files and PDFs, runs shell commands and searches the web, and connects to Google Drive, Gmail, Slack and GitHub ^[3]. Gizmodo, citing VentureBeat, describes a demo in which it posted data analysis into a Slack channel on the user's behalf ^[17]. Data leaves through sanctioned integrations with no frontier model involved and no escalation prompt to approve. Tom's Guide, which credits Perplexity's claim that code and tool use stay inside the local sandbox, said it wanted more detail on exactly what that sandbox can reach and how the permissions are controlled ^[9]. That is the first question any security review asks, and it is currently unanswered.

Then the bill. Gizmodo puts a Mac Mini with an M4 at about $900 and a DGX Spark at roughly five times that ^[15], which is about $4,500 ^[19], one-time rather than metered ^[18]. Add a Pro, Max, Enterprise Pro or Enterprise Max subscription ^[13] and a 24GB VRAM floor that clears most desks out of the running ^[11]. Meanwhile the menu narrows: Computer launched offering 19 models, Portable launches with two ^[16], seventeen fewer ^[20]. Pay more, choose less, and in return know which steps ran where. For a regulated team that has been told no to every agent pilot, that trade may be the only one on offer.

What to watch

Whether enterprise admins can set escalation as central policy, including a hard refusal, rather than leaving each approval to whoever is at the keyboard.
Windows support, due in September, and whether it keeps the 24GB VRAM floor or drops to consumer cards.
Whether Nvidia's reported investment at a $30 billion valuation closes, and whether the hardware requirements loosen if it does not.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence36
Adoption13
Hype gap+31
Incentives74
Confidence57

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Perplexity announced Portable Computer on Tuesday, bringing its agentic Computer platform to Nvidia DGX Spark and other Linux machines with compatible Nvidia graphics cards.
Reported SupportedView cited source
[2]
Perplexity says every task begins on the user's machine, and if the local model reaches a step requiring more advanced reasoning, Portable Computer asks permission before sending that individual step to a frontier model in the cloud; the answer returns to the local workflow while files and other sensitive information can remain on the device.
Reported SupportedView cited source
[3]
Portable Computer can work with files stored on the computer, read PDFs, run shell commands, take dictation and search the web, and users can connect services including Google Drive, Gmail, Slack and GitHub.
Reported SupportedView cited source
[4]
At launch users can choose a 27-billion-parameter Qwen model or Perplexity's post-trained version, PPLX 27B; Nvidia's Nemotron 3.5 Lightning is expected to follow, and users will be able to bring their own model and inference server.
Reported SupportedView cited source
[5]
Gizmodo describes the initial release as limited to Nvidia's DGX Spark desktop supercomputer and Linux machines with Nvidia RTX GPUs.
Reported SupportedView cited source
[6]
Users do not select a privacy mode or write routing rules; the system decides while working, but the user still has to approve the handoff to the cloud.
Reported SupportedView cited source
[7]
Portable Computer packages the local model, inference engine, agent tools, app connectors and an isolated sandbox where it executes code and uses tools.
Reported SupportedView cited source
[8]
Building an agent from a downloadable open model usually means finding and configuring an inference engine, agent harness, sandbox and connectors, and making them work together yourself.
Reported SupportedView cited source
[9]
Tom's Guide reports Perplexity's claim that code and tool use happen inside the local sandbox, but says it would still want more detail about exactly what the agent can access and how those permissions are controlled.
Reported SupportedView cited source
[10]
Perplexity says locally completed work will not consume billing credits, and only the portion sent to the cloud will count toward cloud usage.
Reported SupportedView cited source
[11]
The software launches first on Linux and requires Nvidia hardware with at least 24GB of VRAM, supporting the DGX Spark as well as computers running DGX OS or Ubuntu on ARM or x64.
Reported SupportedView cited source
[12]
Windows support is scheduled to arrive in September, and Perplexity has not announced macOS support.
Reported SupportedView cited source
[13]
Portable Computer will be available to Perplexity Pro, Max, Enterprise Pro and Enterprise Max subscribers, and Tom's Guide notes most people will not have the necessary hardware.
Reported SupportedView cited source
[14]
Computer launched in February as an always-on AI agent that runs autonomously whether or not the user is at the keyboard, running on a dedicated device such as a Mac Mini while the AI processing stayed cloud-based.
Reported SupportedView cited source
[15]
Gizmodo puts a Mac Mini with an M4 chip at about $900 and says a DGX Spark costs about five times that price.
Reported SupportedView cited source
[16]
When Computer launched, Perplexity pitched access to 19 different AI models; Portable Computer launches with two, Alibaba's Qwen 3.8 27B or the Perplexity post-trained PPLX 27B.
Reported SupportedView cited source
[17]
Gizmodo, citing VentureBeat, says Perplexity demonstrated Portable Computer sharing data analysis directly to a Slack channel on the user's behalf.
Reported SupportedView cited source
[18]
Gizmodo notes the hardware is a one-time fee, against the monthly credits bill of cloud processing.
Reported SupportedView cited source
[19]
Five times the roughly $900 Mac Mini price implies a DGX Spark at about $4,500.
Derived Supported
[20]
The local launch line-up is seventeen models smaller than Computer's cloud model menu.
Derived Supported
[21]
Nvidia is reportedly mulling an investment in Perplexity that would value the company at $30 billion, according to Gizmodo.
Reported InsufficientView cited source

Sources

2 independent publishers whose own reporting we read for this story.

gizmodo.com
1 article · August 25, 2026
Perplexity Launches Local AI Model That Will Run on Your GPU Instead of the Cloud
tomsguide.com
1 article · August 25, 2026
Perplexity’s new local-first AI runs on your PC — and asks before using the cloud | Tom's Guide

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder

The build-relevant substance is that Perplexity is shipping a pre-integrated local agent stack: model, inference engine, agent tools, connectors and an isolated code/tool sandbox in one install, against the usual work of assembling an inference engine, harness, sandbox and connectors yourself. Tom's Guide carries almost all of this detail, including the promise of bring-your-own model and inference server and the Linux-first, 24GB-VRAM, DGX OS/Ubuntu-on-ARM-or-x64 support matrix. What builders cannot yet evaluate is the thing that matters most for trust: what the sandboxed agent can reach on the host and how permissions are expressed, which reporting explicitly flags as undisclosed. There are no independent benchmarks of the 27B local models against the frontier cloud fallback, so the quality boundary that triggers escalation is asserted rather than measured.

Operator

Operationally this is a data-residency and approval-workflow story. Tasks are said to start locally, and any step needing stronger reasoning triggers a permission prompt before that single step goes to a cloud frontier model, with files able to remain on device; users get no privacy mode or routing rules to configure, so the escalation policy is the vendor's and the only control point is the per-handoff approval. Cost accounting shifts too: local work reportedly consumes no billing credits, and only cloud portions count. Against that, deployment is gated to Pro, Max, Enterprise Pro and Enterprise Max subscribers on Nvidia hardware with at least 24GB VRAM, Windows is promised for September and macOS is unannounced, and the sandbox's access scope and permission model are not documented, which is exactly the detail a security review would require before connecting Google Drive, Gmail, Slack or GitHub.

Investor

Gizmodo supplies the capital framing: Nvidia is reportedly weighing an investment valuing Perplexity at $30 billion, and the product's exclusive dependence on Nvidia silicon at launch reads as aligned with that. The economics shift consumption from recurring cloud credits toward a one-time hardware purchase, roughly $4,500 for a DGX Spark against about $900 for an M4 Mac Mini, which narrows the addressable base sharply: eligible users need both a paid tier and a high-end Nvidia machine. The local menu is two models versus the 19 pitched for cloud Computer. No revenue, unit-count, or usage disclosure accompanies the launch, and the investment report carries no named primary source, so the funding leg remains unverified.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence36: 0100
Vendor-sourced launch detail, no independent testing
Specifics are concrete and consistent across both outlets (models, VRAM floor, OS support, tiers, roadmap), but every capability and privacy behavior traces to Perplexity's announcement and a company demo. Neither publisher ran the software, no benchmarks compare the 27B local models to the cloud fallback, and the sandbox permission model is explicitly undocumented. The one financial claim is an unattributed rumor.
Adoption13: 0100
Day-one availability, gated and unmeasured
Adoption evidence stops at the release itself, a vendor demo, and a stated billing model. There is no user count, deployment, or usage disclosure, and reach is structurally constrained by a paid tier plus Nvidia hardware with at least 24GB VRAM on Linux only, with Windows still a month out.
Hype gap+31: −100 understated0 aligned+100 overstated
Privacy framing runs ahead of verified controls
The 'runs entirely on hardware you own' framing is overstated relative to what is demonstrated: routing decisions are made by the vendor's system rather than the user, the agent's host access and permission controls are undisclosed, connectors reach into Gmail, Drive, Slack and GitHub, and the local model menu is two options against 19 in the cloud. Tom's Guide partially self-corrects by flagging the missing permission detail, which keeps the gap moderate rather than severe.
Incentives74: 0100
Strong vendor and hardware alignment
All substantive claims originate from a launch announcement and demo by a company seeking relevance. The product runs exclusively on Nvidia silicon at a moment when Nvidia is reported to be weighing an investment valuing Perplexity at $30 billion, and availability is tied to paid Pro/Max/Enterprise tiers, so both parties benefit from the hardware-plus-subscription framing. Gizmodo names this alignment explicitly.
Confidence57: 0100
Consistent reporting, thin verification
Two independent publishers agree on the product facts and no claim in the cluster is contradicted, which supports moderate confidence in what was announced. Confidence is held down by the absence of hands-on testing, any usage data, or primary sourcing for the valuation report, and by the undocumented permission model at the core of the product's value proposition.

Topics and entities

Topics

On-device inference hardware requirements
Local-first AI agents
AI Cost Models: Capex vs Credits
Hybrid Local-Cloud Routing
AI Data Privacy
Agent Sandboxing and Permissions