Home data center with 12 GPUs powers zero-token-cost AI agent development

A Reddit user shared his ambitious home data center build, comprising four separate systems with a combined 12 GPUs. The primary rig uses a Threadripper 3960x (24 cores) with four RTX 3090 Ti cards and 128GB DDR4, drawing nearly 2000W at full load—managed by dual PSUs that have run stable for a month. The second system pairs a 36-core Xeon 8352 with four RTX 5070 Ti cards and 128GB DDR4. The third is an Intel i7-14700k (engineering sample costing $100) with 64GB DDR5 and a single RTX 5090, dedicated to running an embedding model. The fourth system is a Ryzen 5950x with two 5070 Ti cards.

These systems are used for diverse ML workloads. The 3090s are currently training a TTS LoRA distilled from a larger model. The 5070s run Qwen 27B for coding, Nemotron streaming STT for speech-to-text, and Moss TTS for an interactive agent the user is building. He notes that recent Qwen models are good enough to leave running overnight on repo improvements—mostly boilerplate but with real productivity gains and zero token cost. The post highlights the trade-off between hardware investment and unlimited local inference, a growing trend among AI enthusiasts who want full control and no API fees.