In 2025, the famous YouTube gamer and content creator PewDiePie (Felix Kjellberg) stepped into the world of artificial intelligence in a big way. Moving beyond simple cloud usage, PewDiePie built a state-of-the-art local AI lab at home, housing a 10-GPU rig capable of running massive AI language models completely offline. This project not only showcased his passion for technology but also highlighted the growing accessibility of AI hardware setups previously reserved for tech companies and research labs.
This article explores the details behind PewDiePie's multi-GPU AI rig, the hardware and software choices he made, the challenges encountered, and a detailed guide to building your own AI workstation at home.
PewDiePie’s AI Rig: A Local AI Supercomputer
The Hardware Setup
PewDiePie’s home AI system is built around a custom PC housing 10 powerful NVIDIA GPUs:
-
8 modded Chinese RTX 4090 GPUs, each featuring 48GB of VRAM. These are custom-modified versions of NVIDIA’s flagship GPUs, tailored for high-capacity AI workloads.
-
2 NVIDIA RTX 4000 Ada GPUs, adding flexibility and additional memory to the rig.
Combined, these GPUs provide PewDiePie with an estimated total of around 256 GB of GPU VRAM, enough memory to handle large state-of-the-art open-source language models like Qwen 2.5-235B and OpenAI’s GPT-OSS-120B locally. The sheer compute power allows for running models that normally require expensive cloud infrastructure without the delays or privacy concerns of remote servers.
To manage this level of hardware, PewDiePie’s rig features specialized motherboard solutions with split PCIe lanes allowing 10 GPUs to be connected efficiently. Advanced cooling solutions ensure thermal stability in a compact home lab environment. Power delivery is optimized to keep the rig stable under heavy loads.
Software and AI Models
The rig runs multiple large language models locally using frameworks like vLLM. PewDiePie built a custom chat interface called "ChatOS," which connects to an ensemble of AI "bots" that analyze and vote on the best responses, famously called “The Swarm.” These comprise open-source and Chinese-origin models, combining speed, privacy, and accuracy.
By self-hosting locally, PewDiePie eliminates the need for external cloud services, ensuring full data privacy and faster response times. This setup also enables continuous experimentation and customization, as Felix can tweak models and combine their outputs with his own code.
Why Build Locally?
By moving AI workloads off the cloud and onto a personal rig, PewDiePie gains:
-
Privacy: No data is sent to third-party servers, reducing risks of data breaches or surveillance.
-
Speed: Local inference dramatically reduces latency compared to internet round trips.
-
Control: He can run any model, modify parameters freely, and experiment with combining models.
-
Cost Efficiency (long term): Avoids ongoing cloud rental fees though initial hardware investment is high.
Building Your Own Multi-GPU AI Workstation: Step-by-Step Guide
Building a home AI rig like PewDiePie’s is a serious endeavor but increasingly within reach for enthusiasts with the right knowledge and budget. Here is how you can approach building your own multi-GPU AI workstation.
1. Define Your Use Case and Budget
-
Are you building primarily for AI inference (running pretrained models)? Or do you want model training capability as well? Training requires even more GPU power and memory.
-
Budget varies greatly depending on GPU choices. PewDiePie’s 10-GPU rig costs tens of thousands of dollars, but smaller setups with 2-4 GPUs can run many AI tasks well.
2. Selecting GPUs
-
For deep learning and AI, GPU VRAM and CUDA core count are key metrics. Modern cards like NVIDIA RTX 4090, 4080, or 3090 Ti offer excellent AI performance.
-
Multiple GPUs allow distributing AI model workloads.
-
For running large language models, GPUs with 24+ GB VRAM are preferable. PewDiePie’s rig includes GPUs with 48 GB VRAM for extreme cases
-
Buying modded versions (as Felix did) can be cost-effective but requires caution and expertise.
3. Choosing the Motherboard and CPU
-
The motherboard must support multiple GPUs. This means having enough PCIe slots and sufficient PCIe lanes (at least x8 per GPU ideally).
-
CPUs with more PCIe lanes and cores are better — CPUs like AMD Threadripper, Intel Xeon, or high-end desktop CPUs are popular choices.
-
Some motherboards support PCIe bifurcation, letting one slot serve multiple GPUs, but configurations can be complex.
4. RAM, Storage, and Power Supply
-
Aim for plenty of fast RAM — 64GB or more is recommended to support data handling.
-
NVMe SSD storage handles fast data access and large model files.
-
Power supplies must support the combined wattage of all GPUs and components — 1200W or more, often with multiple 12V rails.
-
Proper cooling with liquid cooling or high airflow air cooling is essential to prevent overheating.
5. Assembly and Setup
-
Careful physical installation and wiring is crucial. PCIe risers and GPU brackets may be necessary.
-
Cooling fans and thermal paste should be applied meticulously.
-
BIOS and firmware updates often required to support full GPU setup.
-
Operating system: Linux is popular for AI workstation deployment, especially distributions like Ubuntu.
6. Installing AI Frameworks and Models
-
Install NVIDIA drivers and CUDA toolkits compatible with your GPUs.
-
Popular AI frameworks include PyTorch, TensorFlow, and vLLM for local large language model inference.
-
Download open-source pretrained models (like Meta’s LLaMA, Baidu’s Qwen, or GPT-OSS). Verify licensing and compliance to avoid legal issues.
-
Consider containerization with Docker for easier management.
7. Experimenting and Optimizing
-
Start with smaller models and gradually scale to larger ones as you optimize settings.
-
Multi-GPU workloads require parallelism programming, handled by frameworks or middleware.
-
Monitor thermal and power parameters continuously.
-
Optimize batch sizes, quantization (lower precision), and caching to fit models on available VRAM.
Challenges and Considerations
Building and operating a multi-GPU AI rig is complex and not without issues:
-
High Initial Cost: GPUs, motherboard, CPU, cooling, and power can cost thousands to tens of thousands of dollars.
-
Heat and Noise: Multiple powerful GPUs generate heat and noise, requiring thoughtful cooling and soundproofing.
-
Software Complexity: Setting up drivers, libraries, and models can be challenging for beginners.
-
Power Draw: Such rigs consume significant electricity with consistent high loads.
-
Model Licensing: Running certain large models requires attention to licensing agreements and ethical considerations.
The Future of DIY AI Labs
The explosion of interest in AI democratizes access to advanced technology. PewDiePie’s experiment is not just a personal achievement—it signals a trend toward self-hosted AI models at home, reducing reliance on cloud services and increasing privacy and control.
Smaller multi-GPU rigs now power innovative projects worldwide, from independent research to creative AI-driven art and content creation. As hardware prices decline and software becomes more accessible, building your own AI rig will become increasingly mainstream
PewDiePie’s 10-GPU AI rig likely cost between $30,000 to $50,000 or more.
Here’s a rough breakdown:
-
Each customized RTX 4090 GPU with 48GB VRAM can cost around $3,000 to $4,000 depending on modifications and availability. For 8 GPUs, that’s roughly $24,000 to $32,000.
-
The 2 additional NVIDIA RTX 4000 Ada GPUs could add about $1,000 to $1,500 each, totaling around $2,000 to $3,000.
-
High-end motherboard capable of supporting 10 GPUs, likely specialized for multi-GPU setups, can cost $1,000 to $2,000.
-
A powerful CPU suitable for handling multi-GPU workloads may cost $500 to $2,000 depending on model.
-
RAM (64GB or higher), NVMe SSDs, and other components add around $1,000 to $2,000.
-
Custom cooling and power supply, including high wattage PSUs and liquid cooling, can range from $1,000 to $3,000.
-
Additional costs include PCIe risers, chassis modifications, networking, and assembly.
Factoring all components and customizations, the total easily surpasses $30,000 and could reach closer to $50,000 depending on exact parts and labor.
This estimate aligns with reports describing the rig as one of the most powerful home AI setups, pushing the boundary of consumer hardware capabilities.
Conclusion
PewDiePie’s 10-GPU AI rig reveals how powerful self-hosted AI systems can be built in a home setting with sufficient technical know-how and investment. Running massive language models offline offers speed, privacy, and customization benefits that are hard to match with cloud services.
While his setup may be out of reach for most hobbyists, the principles of GPU selection, motherboard capability, cooling, and software configuration apply broadly to building smaller-scale AI workstations. With the right approach, anyone passionate about AI can follow a similar path to explore, experiment, and innovate right from their home.
