tl;dr: LLM inference costs are dropping ~2.5 orders of magnitude per year, driven by GPU efficiency gains (~2x/2yr), better inference engines (vLLM, MLPerf), MoE architectures, Mamba hybrids that slash local RAM needs, and specialized classifiers like Jev/Laya that are 100x cheaper still. The author predicts LLMs embedded as infrastructure within 1-2 years, frontier-quality models running locally in 3-6 years, and tokens becoming cheaper than tool calls like `grep`—shifting software's value from code to product design, ops, and security.