Reimagining the Data Center: AI Computation as the First Stage of an Integrated Energy System
Year: 2026
Type: preprint
Abstract: Data centers run their chips as cool as possible and purposefully discard the heat of computation. This paper introduces an inverse design. Whereby, AI compute is part of an integrated energy process. An AI facility at an energy site runs silicon-carbide chips deliberately hot. The chips generate heat with geometric operations whose results form a public ground truth for data, accessible by any developer. The generated heat is then used as a secondary energy source. Since every joule a computer draws leaves as heat, two properties decide the worth of a facility: 1) the temperature at which the heat leaves, and 2) the value of the computation that produced it. The proposed design optimizes both. The geometric workload, the spectrum of the covariant Laplace–Beltrami operator, is computed once. This removes redundant computation across AI companies (92% for ten developers) and gives each a ground truth, certified to 0.17%, on which to fit its own weights. Coolant leaves the chips at 300–465 °C; 24.0–35.0% of the heat returns as electricity, and the remainder serves as process heat, with 73% of US industrial demand and 62% of refinery demand within reach. For the same computation and process heat, the integrated system purchases 52–54% less energy than today’s separate systems and avoids 1.6–1.9 Mt of CO2 per gigawatt-year. Standard Silicon cannot participate in this process since its junction limit holds recovery to 3.0%. Each link of the chain is illustrated by an experiment at prototype scale, within a closed dependency chain of physical laws and published benchmarks. The results suggest that a different model for AI infrastructure could result in incremental gains in energy consumption and downstream computational insights.
Author: Swanagon, David
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237366011
Cites:
Cited by:
Related to:
