Four Tiny PCs Chained Together Just Ran a Big Model at Home — and It Didn't Cost a Fortune

Some mornings the AI news feed feels like one long echo. Today is not one of those mornings. Today I'm staring at a $16,000 desk setup that is, technically, four mini PCs glued together with USB4 cables, and I have to admit: the idea is growing on me.

The local-cluster thing that actually made me stop scrolling

GEEKOM took four of their mini PCs, wired them up over USB4, stuffed in 512GB of RAM total, and got DeepSeek-V4-Flash running fully on-prem. The headline number everyone is quoting is 14.61 tokens per second. Let me be honest with you: that's not fast. My brain types faster than that. But that's also missing the point entirely.

The point is you never touched a public cloud server. No API bill, no rate limit, no "we're sorry, this model is at capacity." For anyone who cares about what their data does after it leaves the room — and you should, honestly — that trade is worth thinking about. Enterprise-grade AI without the enterprise lease.

The catch, and there's always a catch: $16,000 is a lot of money for what is still a niche hobby setup. To be fair, that's four machines, not one, and the per-unit math is a lot friendlier than it looks. But you're also running bleeding-edge interconnect config on your own desk, which means you're the first person to discover whatever weird bug happens when your USB4 cable sneezes at 3am.

Real-talk from someone who tinkers: clusters like this are the 2026 version of building your own homelab. You do it because you want to own the stack, not because it's the cheapest path. For most people, renting still wins. For the people who'd never rent a GPU they don't control, this is quietly becoming a real option.

Europe is finally pushing back on smart glasses — and it's a big deal

Meanwhile, across the Atlantic, Germany's digital rights group HateAid just filed a criminal complaint against Meta, Ray-Ban, Oakley, and four major retailers over the Ray-Ban Meta Wayfarer Gen 2. The legal argument is blunt: under Germany's TDDDG data protection act, you can't market a device that looks like an ordinary object but secretly records people. The glasses, HateAid says, are "indistinguishable" from regular sunglasses.

This is where it gets spicy. The complaint names Fielmann, Apollo-Optik, Mister Spex, and MediaMarkt alongside the manufacturers — they want the sales profits forfeited, and executives could face up to two years in prison. In a country where around 41 million people wear glasses, the "there's no place to escape" argument lands harder than it would almost anywhere else.

I'm not going to pretend this is clear-cut. Camera glasses have legitimate uses, and plenty of people happily wear them. But the privacy question was always coming, and Germany just decided to be the one to force it. Watch this one — if Frankfurt's prosecutors take it beyond a preliminary assessment, every wearables maker on the planet is going to re-read their own fine print.

The agent economy crossed an invisible line this year

AWS Marketplace's own numbers tell a story that's hard to argue with. A year ago, "AI agents" ranked 64th among keyword searches on the marketplace. Now it's third. Agent offerings went from roughly 1,000 to more than 4,000 in twelve months, out of almost 40,000 total applications. And the market analysts are projecting the agentic procurement software market to climb from $1.2 billion in 2025 toward $9 billion by 2034.

Here's the part that actually matters, though: the agents aren't just being sold — they're doing the selling legwork. AWS is using agents to handle the due-diligence and admin grind inside its own marketplace, qualifying vendors and shepherding deals. The human sales reps aren't gone; the busywork is.

That's the pattern I keep seeing in every serious deployment. The agents that survive contact with reality aren't the flashy ones. They're the ones quietly doing the boring, multi-step work a human used to do between meetings. Adoption tripled this year, and the reason isn't hype — it's that someone finally measured the ROI and the number made sense.

The water bill is becoming the story

Two smaller items worth a glance. Data center backlash is officially leaking into 2026 election politics — candidates on both sides are suddenly very interested in who's drinking the region's water to cool GPUs. And in the middle of all that, Jason Kelce's Liquid Death ad turned the whole thing into a joke, which is honestly the most effective PR a serious issue has gotten all year.

Quick add-on note: none of this is going to settle the power-and-water fight overnight. But when a water-guzzling data center becomes a campaign issue, that's how you know the infrastructure boom has officially left the tech pages.

---

For my money, the local cluster story is the one to keep an eye on. The pricing will keep dropping, the interconnect will get boring-reliable, and at some point "no cloud needed" stops being a flex and becomes a default. Until then, I'll be over here checking whether my own desk setup can handle 14 tokens per second without catching fire.

If you're curious about the math side of building your own stack — power draw, per-token costs, whether it actually pencils out — I've been keeping a running calculator at Math Calculator that does the heavy lifting.