# HBM demand — X 热门讨论 (2026-09-23 05:11 UTC)
## @SteveSTechMan (Steve) · 09-22 19:19 · ♥40 ↻6 💬0 And if the DGX Spark apocalypse was not enough, an insider update on HBM/LPDDR memory constraints.
HBM is eating the memory industry. Samsung, SK Hynix, and Micron are not “about to flood the market” with cheap LPDDR/DDR, nor is the untested and uncertified (keyword) Chinese market memory.
HBM takes a growing share of DRAM wafer starts (~22% in 2026 → ~30% in 2027) while delivering a much smaller share of bits (~9% → ~13%). HBM4 is roughly 3× the wafer cost of commodity DRAM per bit. That is a structural squeeze, not a blip.
2026 HBM is sold out / locked. 2027 DRAM+HBM allocation is already largely spoken for. SK Hynix’s CEO called 2027 the worst supply year on record and said demand stays above capacity beyond 2030. Chairman Chey put the wafer gap above 20% and the shortage into ~2030. Micron has talked through 2027 with gradual easing in 2028. Samsung’s memory chief has pointed through 2028.
New Korean/US capacity coming 2027–2029 is overwhelmingly HBM- and server-weighted. Long-term AI deals already lock large slices of output into the early 2030s. Supplier inventories have been estimated at under 10 days.
Translation for phones, PCs, and DIY RAM: do not budget for a consumer glut. Expect tight LPDDR/DDR, high ASPs, and product mix shifting upmarket until the next real wafer wave lands, and even then, AI gets first claim.
Buckle up. If you thought it was painful now? Q4 is the on-ramp to 2027, which is going to leave the open community in shock. Your DGX investment just turned into the fastest-appreciating asset you own. https://x.com/SteveSTechMan/status/2102477884065554837
## @antfeedapp (Antfeed) · 09-22 12:58 · ♥30 ↻10 💬1 SAMSUNG CUSTOMERS ARE ASKING FOR MEMORY DEALS AS LONG AS 10 YEARS (DAISHIN SECURITIES)
TIL: Samsung Electronics is seeing customers propose much larger memory purchases than expected as DRAM supply tightens, according to Daishin Securities.
The analyst expects Samsung to secure roughly five years of demand visibility through rolling long-term contracts, with some customers reportedly asking for agreements stretching to 10 years.
Samsung’s 1c DRAM front-end yield is also estimated at around 80%, while HBM sales and pricing are improving.
Daishin now expects Samsung to take the No. 1 HBM market share position next year.
[Get more breaking semiconductor news straight from Korea and around the world, on Antfeed]
#Samsung #HBM #DRAM #Memory #AI #Semiconductors #반도체 $005930 solana:5XqMbyZbtse5UQUzgrYMpeWLWLR6BWFqUm5H3VAWpump $SOXX $DRAM $RAM https://x.com/antfeedapp/status/2102381999940071759
## @SemiconductorsX (Semiconductor Insider) · 09-22 20:00 · ♥37 ↻4 💬4 DRAM and NAND used to move together. Next year they split.
Phones and PCs used to set both tapes. TrendForce now has them on different clocks. DRAM stays tight. Consumer NAND takes the price hit.
Enterprise SSD demand at the big U.S. clouds is expected to run hotter in the fourth quarter than in the third. AI-server eSSD is short.
Lead time has gone from about eight weeks in a balanced market to sixteen. As AI moves from training into inference and agents, the hall needs more live data on flash. Server NAND stays bid.
Client NAND in phones and PCs is the other book. Smartphone and PC recovery is slow. YMTC and other Chinese makers are lifting output. That is where the pressure builds.
DRAM does not get that split. Server and client both look higher. Samsung, SK Hynix and Micron still push wafers into HBM and high-density server DRAM.
Phone and PC DRAM supply shrinks as a side effect. TrendForce expects DRAM demand growth to beat supply growth next year.
AI data centers became the new large buyer. Capacity followed them. The old companion cycle is the casualty. $SKHY $MU $DRAM If eSSD lead times are already 16 weeks and YMTC is adding client bits, does consumer NAND crack first or does the HBM raid keep even phone DRAM tight? https://x.com/SemiconductorsX/status/2102488111548940736