AI inference is obviously profitable (www.seangoedecke.com)

Recent analysis counters the belief that AI inference is unprofitable, suggesting that it can actually be a profitable venture. The evaluation highlights that the cost to run inference using an Nvidia A100 GPU, which consumes 400W of power, can approximate about $1 per million output tokens. With leading AI models, such as OpenAI's GPT-5.4-mini charging $4.50 per million tokens, this indicates a plausible profit margin of 70-80% for inference. Open-source models further reinforce this notion, as evidence from Chinese LLM DeepSeek demonstrates an 80% profit margin while charging less than half of what OpenAI or Anthropic do, showcasing that efficient inference can thrive independently of heavy training investments. The significance of this analysis lies in the implication that the AI inference market can sustain itself even if larger AI labs struggle or fail, as inference providers are not required to shoulder the high costs of developing new training models. Consequently, the robustness of the inference business is positioned as a vital aspect of the AI ecosystem's long-term viability. This insight could reshape the expectations surrounding AI profitability, suggesting that while training models may indeed require high capital, inference alone can remain a viable, profitable domain independent of broader market fluctuations.