Perplexity Just Put an AI Agent on Your PC. Here Is the Real Cost

On September 14, Perplexity released Portable Computer for Windows. It is a version of its Perplexity Computer agent that runs entirely on your own machine. The model, the planner, the scheduler, all of it lives on your PC. Your files never leave the building, and work done locally does not burn any of the cloud credits Perplexity normally charges for agent tasks.

Here is the short answer for a service business owner. Local AI agents are real now, and for one specific kind of operator they are worth it: the one who handles client data they are not allowed to upload, or who runs the same agent tasks so often that cloud credits have become a line item. For everyone else, the cloud is still cheaper, and by a wide margin.

The reason is hardware. Portable Computer needs an NVIDIA RTX GPU with at least 24GB of VRAM. The cheapest new card that qualifies, the RTX 5090, launched at $1,999 and is selling for roughly $4,300 to $4,700 this month because of a memory shortage. That single purchase equals about 19 months of Perplexity Max at $200 per month, or 18 years of Perplexity Pro at $20.

So the question is not "is local AI good?" It is "does my operation have a reason to pay that much up front?" Below is what the tool actually does, what it actually costs, and the test I use to decide.

What did Perplexity actually release this week?

Portable Computer is a local, on-device version of the Perplexity Computer agent, and it now runs on Windows PCs with a qualifying NVIDIA GPU.

Perplexity Computer is the company's agent product. You give it a multistep task, like "go through these 40 invoices and tell me which clients are more than 30 days late," and it plans the steps, runs them, and reports back. Until now, the heavy lifting happened on Perplexity's servers, and each task consumed credits from your plan.

Portable Computer moves that whole stack onto your PC. According to NVIDIA's announcement, the model, the agent harness, the orchestrator, and the scheduler all run on the device. It ships with a local model, Qwen 3.8 27B, that Perplexity post-trained to work with its agent tools, plus its own in-house variant called PPLX 27B. You pick one from a dropdown, click download, and it runs.

It also connects to the things you already use. Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub are supported connectors. So the agent can read your inbox, pull a file from Drive, and post a summary to Slack without any of that content passing through a cloud model.

Two details matter for cost. First, work done locally does not consume Perplexity Computer credits. Second, when a task is too hard for the local model, the agent can ask permission to send that piece up to a cloud model. You decide, per task, whether something leaves the machine.

It is available to Perplexity Pro and Max subscribers on individual and enterprise plans, through the Windows app in the Microsoft Store. Linux and NVIDIA DGX Spark support already existed. This is the release that puts it on a normal office PC.

What does it actually cost to run an AI agent locally in 2026?

The subscription is cheap, the GPU is not, and the GPU is the whole story.

Let me put every number on the table.

The software side is simple. Perplexity Pro is $20 per month. Perplexity Max is $200 per month. Both include Portable Computer. The local model downloads are free. There is no per-task charge for anything that runs on your machine.

The hardware side is where operators get surprised. The requirement is an NVIDIA GeForce RTX or RTX PRO GPU with 24GB or more of VRAM. That rules out almost every laptop and most office desktops. The cards that qualify are the RTX 3090 and 3090 Ti at 24GB, the RTX 4090 at 24GB, the RTX 5090 at 32GB, and the RTX PRO workstation cards.

Here is what those cost right now. The RTX 5090 launched at $1,999 in early 2025. As of September 2026, Tom's Hardware's Newegg tracker put the median price around $4,700, and some retail listings are above $5,000. The cause is a GDDR7 memory shortage as NVIDIA shifts supply toward data center chips. Used RTX 4090s and 3090s are cheaper, but they have also climbed because every hobbyist and small AI shop wants the same 24GB.

Then add the rest of the box. A 5090 pulls up to 575 watts. A used gaming PC will not have the power supply or cooling for it. Realistically you are building or buying a workstation, which puts a complete local AI machine somewhere between $3,500 on the used end and $7,000 or more new.

Now compare that to the cloud. Perplexity Max at $200 per month is $2,400 per year. Claude Max runs $100 to $200 per month. ChatGPT Pro is $200 per month. A service business using one premium agent plan spends about $2,400 a year and never touches a screwdriver.

The break-even on a $4,700 GPU against a $200 per month plan is roughly two years, and that assumes the local model does everything the cloud model would have done. It will not. A 27-billion-parameter model is capable, but the frontier cloud models are far stronger on hard reasoning. The NVIDIA announcement itself notes that Portable Computer will hand off to cloud when a task needs more, which means you may still be paying for the cloud plan anyway.

When does local AI actually make sense for a service business?

Local wins in exactly three situations: data you cannot legally upload, agent tasks you run dozens of times a day, and offices where the internet is the bottleneck.

I want to be fair to the local case, because there is one.

First, restricted data. If you are a bookkeeper, a fractional CFO, a healthcare consultant, or a firm that signs NDAs with confidentiality clauses that name "third-party processors," you may not be allowed to paste client files into a cloud model at all. Many operators just do it anyway and hope. A local agent removes the question. The NVIDIA example for finance is exactly this: point the agent at two years of brokerage statements and tax returns and have it trace the recurring fees, with every number cited to the file and page, and nothing ever reaching a chatbot.

For that operator, the GPU is not an AI expense. It is a compliance expense that also happens to run AI. That is a much easier $4,700 to justify.

Second, high volume. Cloud agent products bill per task or per credit. If your team runs an agent 50 times a day on routine work, like reconciling job tickets against invoices or summarizing every inbound lead email, the credit bill adds up fast. In July I broke down how ChatGPT workspace agents started billing and what that did to monthly costs. Local inference has zero marginal cost. Once the card is paid for, the ten-thousandth task is free.

Third, connectivity. Field service businesses in rural areas, contractors working from job sites, anyone who has watched a cloud agent time out on bad hotel wifi. A local agent does not care.

If none of those three describe you, you are paying $4,700 for privacy you do not need and volume you do not have.

What is the Kept Data Test?

The Kept Data Test is three questions, and you should only buy local AI hardware if you answer yes to at least two of them.

I use this with clients who ask whether they should build a local AI box. It takes five minutes.

One, is there client data you are contractually or legally barred from sending to a cloud provider? Not data you would prefer to keep private. Data where a contract or regulation says no. Check your NDAs and your engagement letters. Most service businesses find the answer is no, and the ones who find yes usually already knew.

Two, would your team run agent tasks more than 30 times a day if they were free? This is the volume question. Be honest. Most operators run a handful of AI tasks a day, and most of those are chat, not agent work. If you are not already hitting credit limits on a cloud plan, you do not have a volume problem.

Three, does the agent's output need to be checked by a person before it goes to a client anyway? If yes, the speed and privacy of local processing matter less, because a human is the bottleneck, not the model. If the agent's work goes straight to a client with no review, a local model that is weaker than the frontier is a risk, not a saving.

Two or more yeses, local is worth pricing out. One or fewer, stay on the cloud plan and revisit in twelve months when the memory shortage eases and the cards come back toward list price.

Why is this landing at the same time small businesses are overbuying AI?

Because adoption is outrunning the plumbing underneath it, and a $4,700 GPU is the most expensive way to find that out.

New Federal Reserve research published this month found nearly 40% of small businesses surveyed were already using or planning to use AI. QuickBooks puts regular AI use at more than three in four US small and midsize businesses. Adoption is not the problem anymore.

Measurement is. Gartner research reported by the Wall Street Journal found 85% of functional leaders plan to increase AI spending in 2026, and 23% of them do not know what return their current AI spend is generating. Almost one in four is buying more of something they cannot measure.

CPA Practice Advisor ran a piece on September 14 quoting agency owner Nick Kurkov, who put it plainly: AI exposes weaknesses that were already there. If customer data lives in four places and nobody knows which spreadsheet is current, a faster tool does not fix that. It runs faster on top of the mess.

A local agent makes this worse, not better, because it is only as useful as the files on the machine it runs on. A cloud agent with a clean CRM will beat a local agent pointed at a Downloads folder every time. Before you price a GPU, answer Kurkov's first question: where does the data actually live? If you cannot name the one system of record for clients, leads, and invoices, that is the project. Not the hardware.

What should a service business do this month?

Keep your cloud plan, clean up your data sources, and put a local box on next year's budget only if the Kept Data Test says so.

Here is the practical sequence.

If you are on Perplexity Pro already and you happen to own a qualifying GPU, turn Portable Computer on. It costs nothing extra and it will save you credits on routine tasks. Try it on one recurring job, like a weekly summary of your Gmail leads folder, and see whether the local model handles it. That is a free experiment.

If you do not own the hardware, do not buy it this month. Prices are at a 135% premium over launch because of a supply problem, not because the cards got better. That kind of premium usually corrects. Run the Kept Data Test, write down your answers, and set a reminder for March.

Either way, spend this month on the boring part. Pick one source of truth for client records. Kill the duplicate spreadsheets. Get your CRM, your calendar, and your inbox talking through whatever automation platform you already pay for, whether that is GoHighLevel at $97 per month or Make.com at $9 to $30. That work pays off no matter where the model runs.

The point of AI in a service business has not changed. It is to give your people back the hours they spend on work that is not the work. A local agent does that for a narrow set of operators with a real data constraint. For everyone else, the cloud plan you already have does it for $200 a month with no shopping trip to Newegg.

Local AI is real, and it is here. Just make sure you are buying it for a reason you can name.

---

If you want a clear picture of what AI can actually do for your specific operation, book a free AI Clarity Call. Thirty minutes, no pitch, you leave with a real answer.

If you want to learn alongside other operators and stay current on what is working, join the Abra AI community. That is where I share what I am actually building.

Subscribe to the newsletter for more breakdowns like this.