The Correctness Tax: Redundancy Energy at Exascale and Beyond

As a computing system grows, its mean time between failures falls as 1/N while long computations do not shorten, so correctness becomes an energy line-item. Using the standard Young/Daly checkpoint model with published-order inputs (per-component failure rate 1000 FIT, sensitivity 200-5000 FIT; 12 GB checkpoint state per component; global checkpoint 600 s; local multilevel checkpoint 60 s), this paper computes system MTBF, the optimal checkpoint interval, and the fraction of total machine energy spent on correctness for N = 1e5 to 1e8 components under three strategies. At 1e5 components (25 MW, comparable to Frontier's published 22 MW) the tax is 16.61% of total energy for global checkpoint/restart and 10.10% for local checkpointing plus algorithm-based fault tolerance at 5% overhead - smaller than folklore claims at current scale. Global checkpointing crosses 50% of total energy at 1.61e6 components; the local strategy crosses 50% at 1.37e7 components; dual modular redundancy costs exactly 50% at every scale and is cheapest beyond those crossings. At 1e8 components the MTBF is 36 s and checkpoint/restart consumes 95.73% (global) or 78.80% (local) of all energy. Model calculations are labelled SIMULATED, NOT MEASURED. The deposit includes the paper PDF and Markdown, the computation script (compute.py) that produces every number in the paper, and its outputs (results.json, results.txt). Computations and drafting were performed with Dave, an AI research system.