The Energy Problem Has a Missing Variable
The Energy Room
Artificial intelligence is the surge driving it.
Most of that energy is spent writing words nobody asked for. Inference, the cost of AI answering, is now over 90% of a model's lifetime energy bill, and verbose output is its largest addressable waste. The industry optimizes the model, the hardware, the prompt. One variable remains unmeasured: the user side. This room documents that variable, and AXIS, the protocol built to convert it into measurable token and energy savings.
The grid math
Inference dominates AI energy cost. Decoding dominates inference. Every unnecessary token is unrecoverable energy.
The Demand Layer
The unnamed user-side variable driving verbose, energy-burning output, and the protocol to test it.
The protocol
A session-governance layer that cuts drift, expansion, and wasted tokens. Black-box evaluable on any frontier model.
The investment frame
Commercial paths, evaluation terms, and current market updates.
The number behind the headline.
TWh stands for terawatt-hour, a unit of electricity. The simplest way to picture it is to build up from your own power bill.
- A kilowatt-hour (kWh) is what your home bill is measured in. Roughly one hour of a microwave.
- A megawatt-hour is a thousand of those. A gigawatt-hour is a thousand of those.
- A terawatt-hour (TWh) is a thousand gigawatt-hours, a billion kilowatt-hours. A country-scale number.
So 1,000+ TWh is a trillion kilowatt-hours, more power than most nations on Earth consume in a year, and that is the load data centers are projected to draw in 2026, with AI the fastest-growing piece of it. The grid was not built for it.
The world spent a decade making the machines faster. Almost no one asked whether the words were worth the watts.
— t r a b o c c o