Every design, an array size and a battery size, lands somewhere on this map: how much of its solar it uses, across, and how busy it keeps the GPUs, up. Colour is profit. The peak sits where the GPUs almost never idle and most of the sunlight is thrown away, because an idle GPU costs far more than a spare panel.
Prototype on the blog's simplified weather model (seeded clear and cloudy days, east-west array, 90% round-trip battery, throttling governor unless On/off is chosen). Solar utilization is the share of the year's generation the GPUs and battery actually use; the rest is curtailed. GPU utilization counts tokens, so a GPU throttled to half power still makes 79% of its tokens (P ∝ f³); the vertical scale is spaced by idle time (each step up is ten times less idling), its top row holds everything at 99.99% or better, and the revenue figures on the right are the same scale in dollars. Shading is profit per MW of GPUs per year as a share of the best design's, finer toward the top; hatching loses money. Faint solid lines hold the battery fixed, in hours of storage at full GPU load; dashed lines hold the array fixed, as a multiple of the GPU load. Lines that reach the 99.99%+ row stop there and are labelled above it. Compute revenue is per kWh the GPUs draw at full power, so it doesn't depend on which GPU: $2.85/kWh is about $2 an hour for a 700 W GPU; profit is revenue minus GPUs, array and battery amortized over the period, per MW of GPUs, before operating costs. Arrays from 1× to 60× the load and batteries from 0 to 48 h are simulated and interpolated between.