Why every question runs the whole model, how much of the work fits a smaller one, and what the default does to the build.
Ranjit Kohli, YYZdata Inc. October 15, 2026. Paper 2 of 5.
The first paper in this series argued that the model is becoming the pipe, and that most of what a business needs does not need the frontier. I wrote then that this last claim had the least measurement behind it of anything in the paper. This paper puts a first number on it.
The same lab sells its largest model at a hundred times the price of its smallest, and builders pick the largest anyway. Each pick is defensible. Added up, it is part of the demand the data center build is meant to serve.
Ask a model what time the delivery arrived, then ask it to review a contract. Both go through the same model, and the model does the same work for every word of either answer. I sent 500 real requests to the smallest and the largest model of one family from one lab. On about half, a grader rated the small model's answer as good as the large one's, at a twenty-fifth of the cost on that sample.
PDF, 7 pages.
Ranjit Kohli is founder and CEO of YYZdata Inc., whose lead product is Signals, an operational intelligence platform.
Paper 1 of 5: The model is the pipe
Paper 2 of 5: The whole model, every time
Paper 3 of 5: Read first, guess last
Paper 4 of 5: Who answers the question
Paper 5 of 5: I watched the last pipe get overbuilt
© 2026 Ranjit Kohli. Published by YYZdata Inc. This paper may be shared and quoted with attribution.