AI is not a better search engine. Using it as one costs more than it returns.

Gartner data projects AI-optimized servers will draw 175 TWh in 2026, up from 95 TWh in 2025. That is an 84 percent year-over-year surge. This year, these specific workloads account for 31 percent of all data-center electricity consumption globally.

A standard search engine finds and ranks what already exists. A generative model produces entirely new text from patterns in its training matrix. The distinction is not subtle. One returns static documents; the other generates active answers.

When we treat the second as the first, we send the entire context window every single time.

The model keeps the whole thread. The human moves forward. What is already obsolete to us stays in the context and gets completely reprocessed. Every new prompt on a bloated, old thread is therefore partly a paid repetition of work already completed.

I have started starting over. When a thread reaches the point where the utility of the information is spent, I kill it and open a new one. Not because the model is underperforming, but because it does exactly what it is programmed to do: it weighs everything that remains.

The question is not whether the model can answer. The question is how much of those 175 TWh is spent holding legacy threads that humans have already left behind.

Who adapts faster—the human or the machine—remains an open question. The electricity bill is not. This is why our 70/30 allocation avoids the over-hyped software layers and continues to accumulate the infrastructure utilities. The market is pricing the application; we are collecting the rent on the waste.