I've got 3x SBCs that can run the Gemma 4 26B MoE on NPU. Around 4W extra power, 3 tokens a second...so that can hammer away at tasks 24/7 without moving the needle on electricity bill
They just use APIs though. There is very little interest within them to do the model engineering and inference in house.