Source: The Register
Google’s new technique for reducing AI model memory consumption has spooked DRAM manufacturers despite representing only a marginal efficiency gain. This exposes how dependent memory vendors have become on the assumption of ever-ballooning model sizes. The real issue isn’t technological—it’s that AI infrastructure costs have become a legitimate procurement bottleneck for cloud providers, forcing them to shop around and negotiate harder rather than simply scale up consumption. Server makers like Dell and HPE are already cushioning guidance and offering vague pricing because they can’t promise customers that memory costs will stay elevated, which means the commodity cycle is finally catching up to the AI hype cycle.