Your Uncle's Frozen Mac Is a Sign You Should Run AI Locally

The Mac That Locked Up

Your uncle called last week because his Mac froze mid-browse, a red warning screen claiming his machine was infected, a countdown ticking down, a phone number to call. It was not a virus. It was a Google ad, one of a wave Ars Technica documented on September 25th, engineered to lock up the browser and scare people into paying for fake support. No malware installed. No files touched. Just a full-screen bluff designed to work because most people do not know the difference between a browser tab and their operating system.

That distinction matters more than it used to, because marketing teams are now running the same browser tabs their uncles are, except theirs are open to ChatGPT with a client's campaign brief pasted in, or a spreadsheet of customer emails dropped into Claude for a quick summary. Every one of those sessions travels through the same ad-supported, occasionally hostile web that just froze a sixty-year-old's Mac over a fake infection warning.

The scareware itself is a nuisance. The bigger question it raises is not. Where does your data actually go when you use an AI tool, what server is it sitting on, and who else gets a look at it along the way? That question has a concrete answer if you are willing to run the model yourself.

What a Local Setup Actually Requires

Running a model yourself does not require a server rack in a closet. For a team of two to ten people, the setup that keeps showing up in hardware guides as of this June is a mini-PC paired with an RTX 3090 with 24GB of VRAM, drawing about $15 a month in electricity. That is the whole server. It sits under a desk and does not ask anyone to think about it again.

The VRAM number is what actually determines what you can run, so it is worth understanding before you buy anything. A 7-8B model at Q4 quantization needs 5-6GB of VRAM. A 13-14B model needs 8-10GB. Pair 24GB of VRAM with 32GB of system RAM, and you are comfortably running quantized 7B-27B models, the range that covers the vast majority of marketing work: drafting, summarizing, rewriting, ad copy variations.

Ollama handles the model management piece. Version 0.34.0, released September 5th, added native support for Gemma 4, gpt-oss, and Llama 4, and it auto-shrinks context to fit whatever VRAM you have, so you are not manually calculating headroom every time you switch models. You install it, pull a model, and point your team at it.

The 70B Myth

Most people assume none of this works unless you are running something huge, a 70B model chewing through 40GB of VRAM on a rack of GPUs you do not own. That assumption is why a lot of teams never try. It is also wrong for the actual job marketing does day to day.

A 7-8B model at Q4 quantization handles drafting, summarizing, and ad copy variations at 30-60 tokens a second on a modest GPU. That is faster than most people type. You do not need the model to write a novel or reason through a legal brief. You need it to take a campaign brief and produce five headline variations, or take a customer email thread and give you the three-sentence version.

If you want more headroom, Qwen3.6-27B at Q4 comes in around 16GB and is currently the highest-scoring model that still fits on a single 24GB card, so the RTX 3090 setup from the last section covers both ends: the fast small model for volume work and a larger one when a task actually calls for more reasoning.

The 70B conversation is a server-farm conversation. Your ad copy is not that conversation. It never was.

Where It Breaks Down

Here is where the mini-PC setup stops being a free lunch. Ollama does not implement PagedAttention, the technique that lets a GPU juggle multiple requests without each one stealing memory from the next. Production notes from this June show P95 latency blowing past one minute once you cross roughly 10 concurrent users. Below that line, the RTX 3090 setup feels instant. Above it, your team starts sending prompts and refreshing the tab, wondering if the server died.

This is a scheduling problem, not a horsepower problem. A single 24GB card can hold Qwen3.6-27B at Q4 and run it well for one or two people typing at once. Ask five people to hit it simultaneously with long marketing briefs and the queue backs up, because Ollama processes what it can and makes everyone else wait their turn rather than intelligently paging memory across requests.

The fix is not more VRAM. It is picking the model size to match your actual concurrent load, not your total headcount. A ten-person team rarely has ten people prompting at the same second. Qwen3.6-27B Q4 at 16GB leaves you enough room on a 24GB card to handle a handful of overlapping requests without falling over, which is the realistic ceiling for a small marketing team's dedicated server.