Chip to Chiller :
The Thermal Stack's Impact on AI Factory Efficiency
New research examining how thermal design decisions influence AI Factory efficiency, Tokens/Watt performance, and infrastructure economics.

New research examining how thermal design decisions influence AI Factory efficiency, Tokens/Watt performance, and infrastructure economics.

AI Factories are increasingly constrained by power availability. As energy becomes the limiting factor for AI infrastructure growth, maximizing Tokens/Watt is emerging as one of the most important measures of AI factory performance.
The output of an AI Factory is Tokens. As a result, AI factories should be optimized to the system-level metric of Tokens/Watt.
This research introduces the Thermal Stack and demonstrates how thermal infrastructure influences GPU die max junction temperature, compute efficiency, and the overall economics of AI Factories.
The findings show that thermal infrastructure is increasingly becoming a performance system rather than simply support infrastructure.
A well-engineered Thermal Stack can improve AI factory efficiency by more than 30%.
Cooler GPUs generate more useful compute per watt.
GPU package design can significantly influence efficiency.
Advanced coldplate designs can materially improve AI factory performance.
Power—not compute—is becoming the primary constraint on AI Factory growth.
Explore the complete analysis, engineering tradeoffs, efficiency calculations, and recommendations behind the Thermal Stack framework.