Why Falling AI Token Prices Don’t Mean Lower Costs

For decades, Moore’s Law shaped how we think about technology costs. Faster chips meant lower prices over time. More power, less expense. That pattern trained leaders to expect efficiency gains to translate directly into savings. In artificial intelligence, the story sounds similar at first. The cost per token for large language model inference continues to fall. According to Epoch AI, token pricing has dropped sharply in recent years. At the unit level, AI is getting cheaper. Yet in real-world systems, total spending is rising. As a cloud engineer working with AI workloads, I see this disconnect daily. The per-token price may decline, but the number of tokens consumed per task is growing at a much faster rate. The result is a cost illusion. On paper, inference looks inexpensive. In practice, total AI spend often increases. Let us unpack what is really happening. The cost illusion: cheaper tokens, higher bills Research from Andreessen Horowitz and Epoch AI shows that LLM inference cost