Grey blocks = model weightsGreen liquid = KV cacheFlat plate = 1 GB runtimeLower tray = system RAM overflowRed lamp = won't loadNeedle = decode tok/sSeven glass tanks, one per device. Grey weight blocks float on green KV-cache liquid; overflow falls into a slow RAM tray. A red lamp means it won't load.Memory needed = weights (parameters × bits ÷ 8) + KV cache (2 × layers × KV heads × head size × bytes × context × users) + about 1 GB of runtime. Each tank is as tall as its device's memory and receives the same workload. The scale is linear up to 32 GB, then compressed four times above the zigzag marked ÷4, so a 16 GB card and a 128 GB DGX Spark both stay readable and one workload stands at the same height in every tank. The runtime plate and the KV cache stay in fast memory; when the total passes a tank's capacity line, weight blocks are pushed over the weir and down the glass downpipe into the drip tray: on a discrete card the runtime keeps those layers in system RAM (dual-channel DDR5, 96 GB at 89.6 GB/s) and every token waits on them. If even the tray would overflow, or on unified memory (DGX Spark, the Mac), where the weir is blanked off because there is no slower tier, the gate shuts and the model does not load. Decode speed assumes 70% of peak bandwidth, re-reading the active weights plus every user's KV cache for each token.