While Meta Puts Robots in Its Data Centers, Your Marketing Team Can Run Its Own AI at a Desk

Robots Reseating Servers, Meanwhile at a Marketing Agency

Meta is putting robots inside its own data centers. WIRED reported in late August that the company is testing hardware from Kinova, Watney Robotics, and ABB to handle cable swaps, server resets, and reseating failed components across campuses like Altoona and Prometheus. The goal is cutting technician workload, in some cases by up to 80%, because the infrastructure spend behind AI right now is so large that even the physical labor of running the servers needs automating.

That is the scale Meta operates at. Robots walking the aisles of a data center so a human does not have to.

Your marketing team is not doing that. But the underlying instinct, run more of this yourself instead of paying someone else by the seat, applies just as much to a five-person shop as it does to a hyperscaler. The difference is you are not buying robots. You are buying a workstation with a decent GPU, or using the Mac already on someone's desk, and pointing it at a model that fits.

This is not a theoretical option anymore. Ollama shipped version 0.33.1 in late August, adding support for newer models and improving performance on Apple Silicon. A small team can install it with one command, load a model sized to their hardware, and start drafting content or running analysis without sending client data to a cloud vendor. The tools that used to require a data center now run at a desk.

The Hardware That Actually Works

The setup that actually holds up for a team of two to ten is boring on purpose. A dedicated workstation with an RTX 3090 or 4090, 24GB of VRAM, sitting in a closet or under a desk. Or a Mac with 32GB or more of unified memory, since Ollama's MLX backend improvements this year made Apple Silicon a legitimate option instead of a slow fallback.

VRAM is the number that matters, not the number of cores or the marketing on the box. A 7B or 8B model at Q4 quantization needs 4 to 6GB and runs 30 to 60 tokens per second once it's on GPU. A 27B model, which handles most content drafting and analysis work a marketing team throws at it, fits comfortably on a 24GB card. Go bigger than your VRAM and the model spills onto the CPU, and everything slows to a crawl. This is the part people skip reading about, then wonder why their setup feels sluggish.

Cost-wise, June 2026 guidance put a dedicated workstation or mini-PC at roughly $15 to $25 a month in electricity. That's the operating cost, not the hardware itself, which you buy once. Compare that to $500 a month for ChatGPT Teams and the math starts to explain why teams are looking at this seriously instead of treating it as a hobbyist project.

Picking a Model That Fits Your Card

Once you know your VRAM budget, picking the model is the easy part. Ollama's August update added support for Llama 4 Scout and Qwen, and both run well within the ranges we just covered. Scout at 7-8B quantized is the workhorse for daily writing tasks, social captions, first drafts, email variations, and it does that at 30 to 60 tokens per second on a card with 4 to 6GB free. Qwen scales up cleanly to 27B if your team needs more nuance for longer-form content or research summaries, and a 24GB card handles that without spilling to CPU.

Tokens per second sounds like a spec sheet number until you translate it. Thirty tokens per second is roughly conversational reading speed, which means a draft blog post generates about as fast as you could read it out loud. Sixty tokens per second feels closer to instant. Below 10, you're watching the model think, and a content team burns real time waiting instead of editing. This is the number that decides whether a setup gets used daily or gets abandoned after a week.

One thing to watch: Ollama defaults to a 4k context window, which silently truncates longer marketing documents, briefs, style guides, whatever you feed it. If your team pastes in a 6,000-word brand doc and the output ignores half of it, that's the default talking, not the model failing you.

Where This Setup Breaks

The one-line curl command that installs Ollama takes about ninety seconds. That part genuinely works exactly as advertised. What it does not do is warn you when your model has quietly outgrown your card.

Load a 27B model on a GPU with 16GB of VRAM instead of 24, and Ollama does not fail loudly. It spills the overflow onto CPU and RAM, and your 40 tokens per second becomes something closer to 3 or 4. Nobody on your team files a bug report for this. They just stop using the tool because it feels broken, and they go back to pasting things into a browser tab.

The context window causes a quieter version of the same problem. Ollama defaults to 4k tokens, which is fine for a caption or a short email but nowhere near enough for a real brand style guide or a client brief with six months of notes attached. Feed it a long document past that limit and it does not error out. It just drops the earlier part of the text and answers based on whatever fit. The output looks confident. It is working from half your document.

Installing the model is a command. Running it well is a sizing exercise, and most of the failures people blame on the model are actually failures of matching hardware to workload.