Ran an SDUI stack for two years and the LLM part was never the bottleneck, it was the allowlist of what a payload could actually call.