AI Computing · Watching along
What's PewDiePie up to? The world's biggest YouTuber just built a mining rig
Updated 20 July 2026 · about a 7 minute read
I put PewDiePie's computer build video on expecting background noise, and about six minutes in I sat up. Aluminium extrusion frame. Open air, no case. Two PSUs mounted side by side. GPUs staggered one-up one-down for airflow, and a wall-power problem serious enough that he called an electrician. I've built this machine. It's sitting in my garage. Mine mined Ethereum.
If you're reading a mining site, there's a fair chance you know exactly what I mean already. But let's do this properly.
Who he is, if you've somehow avoided knowing
Felix Kjellberg - PewDiePie - is the biggest individual YouTuber there has ever been. A Swede who started posting game commentary around 2010, became the most-subscribed channel on the platform for years (about 110 million subscribers), moved to Japan, and then did something almost nobody at that level does: stepped back. The daily-upload machine stopped, and what's emerged since is something much more interesting - a bloke in his thirties doing deeply nerdy projects at his own pace. He built his first gaming PC. That led to Linux, then to writing his first code, then to degoogling his life and self-hosting everything he could. His words at the top of the build video: in order to ascend, he needs more compute.
Why do we care? Because when the largest creator on the platform spends a video showing tens of millions of people that running your own AI on your own hardware is a joyful, doable hobby - that moves the culture more than every "best GPU for LLM" article ever written. Mine included.
What he actually built
The machine's called Odysseus, and the spec (from the build video) goes like this: a Threadripper 7975WX on a Pro WS motherboard, 96GB of RAM ("only half as much as I need"), two 1300W PSUs, and eight RTX 4000 Ada cards at 20GB each - 160GB of pooled VRAM. Not glamorous cards, as he says himself. But get eight of them and it gets interesting.
The case is the bit I love. He couldn't buy what he wanted, so he built a frame from 3030 aluminium extrusion, tapped his own threads (badly at first, by his own admission), made his own GPU brackets, and rebuilt the layout four times because he kept finding better arrangements. He mounted the cards staggered - one low, one high - so they could breathe in open air.
The build video. The frame construction starts about a third of the way in.
He built a mining rig. He may not know it
Go down his problem list and tell me you don't recognise it. An open-air frame because cases can't hold the cards - that's every mining frame ever sold. Aluminium extrusion construction - mine were 2020 profile, his is 3030, same t-nuts, same fiddly little markers. Dual PSUs because one can't feed the GPUs. Card spacing chosen for airflow rather than slot order. A power bill conversation with a partner, handled with a prepared script (his "it pays for itself" maths for Marzia is, word for word, the speech every 2021 miner gave). An electrician visit. Dreams of solar. Even his scripted justification - $2 per GPU-hour in the cloud, times seven cards, times eight hours a day - is the same buy-vs-rent arithmetic we all did, and it's about as honest as ours was.
The one problem he had that miners never did is the interesting one: he needed the cards to work together. A mining rig is eight independent workers who happen to share a house - if one card dies, the other seven keep hashing. Tensor parallelism (splitting one model across cards so they act as a single brain) is a different game, and it's where his build gets properly educational. He discovered the hard way that vLLM-style tensor parallelism wants a power of two - six GPUs won't do, it's four or eight - and rather than sell down to four like a sensible person, he bought two more - gambler's logic that anyone who kept buying GPUs into 2022 will recognise from the inside.
Then his motherboard wouldn't do x8x8 bifurcation (splitting one PCIe slot to feed two cards), ASUS didn't respond, and he ended up flashing a BIOS sent to him by a random forum stranger who "seemed like a good person". It worked. Eight cards, TP8, one brain. The kindness of randoms is real infrastructure in this hobby and always has been.
What he's running on it
This is where it stops being a build story. Per Tom's Hardware's coverage, he's gone all-in on the self-hosting side since: running Qwen models locally, building his own chat interface, bolting a deep-research function onto it, and - this is the bit I'd have watched a whole video of - running a council of chatbots that answer in parallel and vote, pitting models against each other to get better answers. He's talked about training his own model next. The line Tom's pulled out as the standfirst is the purest hobbyist sentence I've read all year: "I like running AI more than using AI."
I understand that sentence completely. It's why there's a benched vLLM stable on my own 48GB 4090 rather than a ChatGPT subscription, and why I spent a month changing one flag at a time and re-running benchmarks. I'm not sure I could defend the hours on any practical grounds, and I don't feel much need to.
What the rest of us can take from it
A few things, some practical, one bigger.
Practical first. His eight-cards-of-20GB approach against my one-card-of-48GB approach is a genuine fork in the road for anyone building local AI, and neither of us is obviously right. Pooling many cheap cards gets you VRAM headline numbers, but tensor parallelism over PCIe taxes every token, wants powers of two, and multiplies your points of failure - I went the other way partly because one big card avoids the whole circus, and I'm only now (nervously) heading to two. His 150W idle observation surprised him; it wouldn't surprise anyone who's run a GPU rig 24/7 and metered it. And his bifurcation saga is a warning worth having before you max out a consumer board: the workstation-class motherboard tax exists for a reason, and the workaround market is sketchy BIOS files from forum strangers.
The bigger thing. For years, mining-rig knowledge looked like a dead skill - frames, risers, multi-PSU wiring, airflow spacing, wall-circuit maths. Then local AI arrived and it turned out the miners had been running the dress rehearsal. The hardware culture transferred almost completely; only the workload changed. PewDiePie arrived at the same machine from the opposite direction, by pure first-principles stubbornness, which I find weirdly reassuring - the form factor isn't a mining thing or an AI thing, it's just what a home-scale GPU farm wants to be.
His rig is one flight sheet away from being a miner, and my old miner is now doing his rig's job. I keep turning that over and I'm not sure which direction of travel is the funnier one.
What I'm watching for now: whether the own-model plan survives contact with what training actually costs, and whether the council idea - many small models voting - beats one big model at anything measurable. If he publishes numbers, I'll be first in the comments. For now, I'm just enjoying watching 110 million people learn what a t-nut is.