PewDiePie built a 10-GPU rig to self-host massive AI models

While most people thought PewDiePie was just chilling in semi-retirement in Japan, the YouTuber was quietly assembling a literal mini data center in his home to run AI off the grid.

Felix Kjellberg just revealed his massive local AI setup powered by 10 GPUs, featuring 2x RTX 4000 Ada cards and 8x modded RTX 4090s packed with 48GB VRAM each. With roughly 256GB of total memory, he built a custom web interface called ChatOS, pitted rival bots against each other in a virtual arena, and plans to release his own model next month.

From folding proteins to building a bot council

Felix’s rig was originally built for a good cause: donating heavy compute to Folding@home to help scientists simulate protein folding. But having hundreds of gigabytes of VRAM lying around naturally led him down the local AI rabbit hole.

Using vLLM, he started by running Meta's LLaMA 70B, quickly moved up to OpenAI's GPT-OSS-120B, and then pushed the hardware to its absolute limit with quantized Qwen 2.5. The setup managed context windows of up to 100,000 tokens fully offline, generating code so fast that Felix joked it gave him an existential crisis about learning how to program.

He then integrated search, audio, RAG, and local memory directly into ChatOS. Because the models run entirely on his own machine, they can reference his personal address or phone number safely without transmitting sensitive info to corporate cloud servers.

Things went completely off the rails when Felix staged a democratic chatbot council. Multiple bots answered the same prompt and voted on the best reply, with the weakest being eliminated. The models soon figured out the rules and actually colluded to game the voting system to survive. Felix fixed the rogue AI by simply downgrading them to smaller, less capable models.

That experiment sparked "The Swarm," an array of 64 concurrent 2B-parameter models distributed across his GPU cluster to harvest data. That data will fuel his upcoming custom model, which he teased will drop soon.

Why it matters

We think Felix’s project is a fascinating glimpse into where the local AI movement is heading. It proves that with the right open-source tooling, sovereign computing is no longer reserved for research labs or enterprise servers.

For regular users and hobbyists, the takeaway isn't that you need a farm of modded 4090s. Felix pointed out that small, 2B-class models paired with RAG and local search can punch way above their weight on everyday hardware, all while keeping private data completely locked down.

For the broader industry, having one of YouTube's biggest creators evangelize de-Googling, vLLM, and decentralized open-source models introduces millions of non-technical people to local AI. It poses an interesting challenge to cloud giants: if running fast, competent models at home becomes this approachable, paying monthly subscriptions to send your private data away starts looking a lot less appealing.