This position paper argues that Large Language Models (LLMs) can support more energy-efficient software development by automati- cally converting high-level interpreted code into more performance- oriented compiled implementations. We examine this claim using Python solutions to LeetCode problems and their C++ counterparts generated by an LLM. While functional correctness is preserved in most cases (87% acceptance), the transpiled C++ programs yield substantial efficiency gains, reducing runtime and total energy con- sumption by around 94%. Notably, these improvements hold for problems both before and after the model’s knowledge cut-off, sug- gesting that the gains arise from properties of the compiled code rather than memorisation of known solutions.