Stop pretending self-hosting is cheaper. It's not. We do it for different reasons and we should say so.
A Reddit user argues that self-hosting large language models is not economically cheaper than cloud-based solutions. They calculated that their personal rig, costing around $2800 and consuming significant electricity, incurs a higher per-token cost than renting cloud GPUs like an H100. The user concludes that the primary motivations for self-hosting are privacy, control, and the desire to tinker, rather than cost savings. AI
IMPACT Self-hosting LLMs is often perceived as more cost-effective, but this analysis suggests privacy and control are the main drivers, not economics.