Yes, I've worked with the Ubuntu VM AI development package a few times. From what I've seen, Ollama running inside the Ubuntu VM doesn't directly utilize the Apple Silicon GPU -- mainly because the virtualized environment can't access Apple's Metal API. The GPU acceleration is limited to what the VM layer can expose, so most of the heavy lifting still happens on the CPU side.
If you're looking for better GPU performance on Apple Silicon for AI workloads, I'd suggest running Ollama natively on macOS instead of inside a VM. That setup allows direct GPU access, and you'll notice a big speed improvement for model inference.
I ran into the same issue when testing local LLM setups for client AI apps at Guru Technolabs, where we handle end-to-end AI development and optimization. Running Ollama natively has consistently given us the best performance results.