GS on memory ( $MU / $SNDK ): "The key demand driver of both the 2017-2018 upcycle and the current upcycle is strong server demand. While the 2017-2018 upcycle was driven by cloud data center expansion, the current cycle is driven by massive AI server build out...memory demand is now closely tied to server demand where server segment comprises around 50% of DRAM demand and ~40% of NAND demand." "given the long lead time (takes at least ~3 years to construct a fab and produce output), we expect conventional DRAM capacity addition pace by 2030 to be slower than the 2017-2018 upcycle." "Based on our checks, LTAs are being discussed to include sizable prepayments and meaningful penalty clauses for non-compliance which could guarantee stronger binding power. Pricing is being discussed within a pre-determined range with a floor price that guarantees a high margin level on a sustainable basis." Expect supply shortage to persist at least until 2028 and next year to be tighter than this year "We now expect 2026E/2027E/2028E DRAM S/D to be in a larger undersupply of 5.0%/5.9%/3.9% (vs. prior expectations of 4.9%/2.5% undersupply in 2026E/2027E)" "We now expect 2026E/2027E global DRAM demand to grow 28%/20% yoy (vs. +25%/+17% yoy previously), and introduce 2028E estimate of +19% yoy growth." "We now expect 2026E/2027E/2028E NAND S/D to be in a larger undersupply of 4.4%/4.6%/3.0% (vs. prior expectations of 4.2%/2.1% undersupply in 2026E/2027E)" "We now expect 2026E/2027E global NAND demand to grow 23%/20% yoy (vs. +22%/+15% yoy previously), and introduce our 2028E estimate of +19% yoy growth." HBM: Expect a tighter 2027 S/D, and a significant ASP increase to lead to a +100%+ growth in TAM to US$116bn We now expect 2026E/2027E/2028E HBM S/D to be in a larger undersupply of 5.4%/6.0%/4.3% (vs. prior expectations of 5.1%/4.0% undersupply in 2026E/2027E) Significantly raise HBM ASP forecast as we expect a catch-up to conventional DRAM Raise 2026E/2027E/2028E HBM TAM to US$56bn/US$116bn/US$168bn