The planner assigns ten percent of twenty-eight hundred billion parameters to shared weights, giving two hundred eighty billion. The remaining ninety percent is split among sixty-four experts. Activating eight experts adds three hundred fifteen billion, so shared plus active experts total five hundred ninety-five billion. Both bars use point eight pixels per billion parameters. This is the written mixture model, not a verified architecture specification. At half a byte per parameter, all twenty-eight hundred billion parameters need fourteen hundred decimal gigabytes of weights. The shared plus selected experts need two hundred ninety-seven point five. Both bars use point three five pixels per decimal gigabyte. The difference is inactive expert weight storage. Choosing fewer active experts changes this working set, but does not make the complete checkpoint disappear or establish a measured performance result. The active weights require two hundred ninety-seven point five gigabytes, while entered VRAM is ninety-six. The source clips weight placement to ninety-six, leaving no space for its sixteen-gigabyte cache. Both bars use one point six pixels per assigned gigabyte. Its reported RAM counts inactive weights but omits the active overflow and spilled cache. The default point-five token rate is an assigned disk-swap branch. These omissions mean the page is a topology heuristic, not safe provisioning or a throughput benchmark.