Google Research released TurboQuant last week. It is a software-based AI memory compression technology that makes inference eight times faster and increases capacity six times without losing accuracy. Memory companies like Micron sold off because the market thinks this will lower demand for expensive high-bandwidth memory. The theory is that inference could shift to cheaper DDR5 memory running on CPUs. The article argues this might be an overreaction. When DeepSeek introduced similar efficiency gains, AI companies just used the savings to build bigger models. Making AI cheaper to run usually increases overall adoption, so high-bandwidth memory demand for training will probably keep growing.