TEMPO: A Chunk-Level Workload Memoization Architecture for Thermal and Reversible-Computation Reuse in AI Compute Clusters

Thermal management in AI compute clusters is dominated by two established but largely separate research traditions: predictive, telemetry-driven scheduling that reacts to or forecasts heat from live workload signals, and materials-level bio-inspired cooling actuators. This paper does not propose a new physical cooling mechanism or a new control-theoretic technique in isolation; both are already well developed in the literature reviewed in Section 3. Instead, it proposes a narrower architectural component, termed TEMPO (Thermal-Evidence Memoized Path Orchestration), which applies content-defined chunking and deduplication — a mature storage-engineering technique — to index and reuse thermal behavior profiles and, where structurally applicable, reversible-computation ("uncompute") paths for recurring sub-structures of AI compute graphs. The central premise is that a substantial share of production AI compute recurs at the sub-graph level even when whole jobs differ, and that this recurrence is not currently exploited by either the predictive-scheduling or the reversible-computing literature. This paper positions the proposal relative to prior art across five adjacent fields, states explicitly what it does not claim, and proposes a minimal empirical validation plan. No prototype, simulation, or professional patent-clearance search has been conducted; this is a design proposal offered for review, critique, and replication, not a report of results.