Trade Whisperer(@TradexWhisperer):$MU $SKHY AI is Memory Bound. Period. At SEMICON Taiwan 2026, $GOOGL Cloud's Nikhil Cherian said high performance memory now makes up over 75% of an AI server's hardware bill of materials. AI is shifting from compute bound to MEMORY BOUND as multimodal and MoE architectures scale. Google's response is a dual track approach. The inference focused TPU 8i packs 288GB of HBM with 384MiB of on chip SRAM to keep KV cache local. The training focused TPU 8t links 9,600 chips into a shared HBM pool of up to 2PB. $DRAM and NAND combined are projected to hit 47% of CSP capex in 2026, rising to 68% in 2027. Server DRAM contract prices are expected to surge roughly 270% this year.

2026.09.02 16:32

$MU $SKHY AI is Memory Bound. Period. At SEMICON Taiwan 2026, $GOOGL Cloud's Nikhil Cherian said high performance memory now makes up over 75% of an AI server's hardware bill of materials. AI is shifting from compute bound to MEMORY BOUND as multimodal and MoE architectures scale. Google's response is a dual track approach. The inference focused TPU 8i packs 288GB of HBM with 384MiB of on chip SRAM to keep KV cache local. The training focused TPU 8t links 9,600 chips into a shared HBM pool of up to 2PB. $DRAM and NAND combined are projected to hit 47% of CSP capex in 2026, rising to 68% in 2027. Server DRAM contract prices are expected to surge roughly 270% this year.