NVIDIA research revealed that 40-70% of the queries of LLMs can be answered reliably with a specialized small language model (SLM). This microservice method also enhances observability, simplifies debugging, and provides predictable scaling. With the growing adoption of AI in enterprises, many have found that large models are not the most viable choice for all enterprise workloads.