Simulation wall-clock t
sim
8 h
CPU cores n
cpu
64
CPU price p
cpu
$0.04/core·h
Training simulations D
2,000
Training GPU-hours
384
GPU price p
gpu
$2.0/GPU·h
Inference time t
inf
2 s
Cost per simulation c
sim
$20.5
Data generation C
data
$41.0k
Training C
train
$768
≈ 37 sim-eq
Break-even N*
2,037 queries
Traditional solver
AI surrogate
Solver cost grows linearly with queries; the surrogate has a fixed upfront cost and a nearly flat line. The crossing point is the break-even query count.
Beyond N*, every additional query saves ≈ c
sim
. Savings shown at 5× the break-even volume:
$330k
.