The LLM inference itself may be more efficient (though this may be impacted by different throughput vs. latency tradeoffs; local inference makes it easier to run with higher latency) but making the hardware is not. The cost for datacenter-class hardware is orders of magnitude higher, and repurposing existing hardware is a real gain in efficiency.