A friend texted me the other night, a little sheepish. He had just seen a scary headline about AI drinking up the world's water, and he wanted a straight answer to a simple question: every time he asks a chatbot to clean up his emails, is he quietly hurting the planet? He was half-ready to feel guilty about it.

The honest answer is smaller than he feared and bigger than he thinks, both at the same time. One answer costs almost nothing. A billion answers a day is a different story. This finale is me handing him, and you, the actual receipt for a single AI reply.

We spent Part 1 on the chip, quartz from a North Carolina mine turned into the most energy-dense silicon humans make. We spent Part 2 on the rack, a single cabinet that now pulls more power than your street, stacked by the thousand into gigawatt buildings. Both times the machine had not answered a single question yet.

Now it does. You type a question, hit enter, and a few seconds later words appear. This is the moment all that steel and silicon and water was built for. So here is the question this whole series was walking toward: what does that one answer actually cost?

AI does not read words, it reads tokens. A token is a chunk of text, very roughly three quarters of a word. "Silicon" might be one token, "unbelievable" might be three. When you send a prompt and get a reply, you are paying for two piles of tokens: the ones you put in, and the ones the model writes back. The output pile is always the expensive one, because generating text takes far more computing than reading it.

So the true unit price of AI is not "per question." It is per thousand tokens. Everything else is arithmetic on top of that.

Here is what the leading models charge to process one million tokens, as of mid-2026. I have converted each to the price of 1,000 output tokens, the pile you actually pay for, so you can compare them like a menu.