In a recent podcast, Google's AI Edge tech lead Cormac Brick argued that the key constraint for on-device AI is memory cost, not compute power, noting that a Raspberry Pi with 6 gigabytes of RAM now costs two-and-a-half times what it did at launch. Brick stated that phone manufacturers are shipping less DRAM in their 2026 devices than the year before, and that DRAM costs remain stubbornly high relative to the rest of the silicon stack. He advocates for using radically smaller models, defining "small models" as those with 1 to 4 billion parameters, which ship inside Android AI Core and Apple Intelligence. However, these models require 4 to 8 gigabytes of DRAM to host comfortably, restricting them to laptops, mobile phones, and higher-end electronics, and putting them out of reach of lower-tier web browsers and the wider IoT market. As an example, Google's Gemma 2B model has its weight memory driven down to 841 megabytes through quantization averaging 2.9 bits per weight, but the practical floor for deployment is roughly 4 gigabytes; on a device with 2 gigabytes of RAM, it simply won't run.