HBM4

Headlines · 7

1. 3h
   Morgan Stanley lifts Q4 DRAM, NAND contract price forecasts; compute still top AI pick

   A Morgan Stanley report raised its memory contract-price forecasts: Q4 2026 DRAM contract prices are now seen up 10–15% quarter over quarter (from +3–8%), with PC DRAM at +13–18% and server DRAM at +10–15%, while NAND moves to +15–20% from +0–5% and enterprise SSDs reach +23–28%. The bank forecasts 2027 HBM average selling prices up 121% year over year, with HBM4 at US$3–5 per Gb and HBM4e above US$5, but says Nvidia's Rubin ramp is behind schedule, with 2026 VR200 NVL72 rack shipments estimated below 10,000 units versus an original target of about 15,000. It keeps its AI supply-chain priority of compute first, networking second and selective memory allocation, and projects DRAM capex of US$98 billion in 2026 (+82%) rising to US$150–160 billion in 2027.

   科技新报 · 🔥 9
2. 2d
   NVIDIA Rubin said to stick with 12-layer HBM4 as HBM prices seen up 121% in 2027

   A supply-chain survey cited by Counterpoint Research indicates most of NVIDIA's next-generation Rubin AI accelerators will keep the originally planned 12-layer HBM4 stacks, rather than shifting broadly to 8-layer HBM4 as earlier rumors suggested. NVIDIA's published roadmap lists 12-high HBM4 with up to 288GB of memory per GPU and bandwidth up to 22TB/s; Korean media had reported that NVIDIA asked Samsung Electronics and SK hynix for more 8-layer HBM4 supply, and industry observers say the two accounts need not conflict — some products may use 8-layer stacks to ease capacity pressure while flagship Rubin parts stay at 12 layers. Separately, TrendForce forecasts HBM average selling prices could rise 121% year over year by 2027, driven mainly by the mix shift from HBM3E to pricier HBM4 rather than pure shortage; because the costly base die is spread over fewer memory layers, 8-layer HBM could cost 10% to 20% more per Gb than 12-layer in 2027.

   cnBeta · 🔥 3
3. 3d
   CXMT's 4F² DRAM breakthrough triggers global memory chip selloff

   On Oct. 10 CXMT president Cao Kanyu said at the 4th Integrated Chip and Chiplet Conference that the company had made key progress on a 4F² DRAM architecture built around its self-developed VSG vertical-surrounding-gate transistor, and expects to ship next-generation DDR5 RDIMM products combining 4F² with wafer-level hybrid bonding by year-end. Memory stocks then fell hard: Samsung closed down 7.39%, SK Hynix dropped 10.09%, the KOSPI lost 5.46% and triggered a SIDECAR program-trading halt, and Micron slipped below $1,000 a share, while CNBC tied the wider Asia tech rout partly to Fed rate-hike fears and worries about AI investment returns. The report also says CXMT posted first-half revenue of 150.3 billion yuan, up 873.64% year on year, with net profit of 77.6 billion yuan and a global DRAM revenue share that rose from 4% in Q2 2025 to 10% in Q2 2026, though its HBM3E is not expected to reach mass production until 2027, roughly three to four years behind Samsung and SK Hynix.

   36氪 人工智能 · 🔥 0
4. 3d
   Ex-Intel CEO Gelsinger calls HBM 'bad memory'

   Former Intel CEO Pat Gelsinger, in an interview with a16z, again criticized high-bandwidth memory as "bad memory," citing low bit density, limited bandwidth and power/thermal problems, and calling it "catastrophic" for AI workloads that tightly couple memory and compute, according to an Oct 14 report that cites a Wccftech post from Oct 13. The report notes HBM3/3E uses a 1,024-bit bus while HBM4 moves to 2,048 bits, requiring many parallel connections and microbumps, and cites an analysis by KAIST finding that once HBM power exceeds roughly 30W, heat from the PHY interface between the GPU or ASIC and the HBM stack itself becomes a main source of thermal stress. These are Gelsinger's and third-party views; no new product, benchmark or roadmap was announced.

   IT之家 AI · 🔥 0
5. 5d
   HBM yield ceiling: TrendForce forecasts 2027 supply tightness, blended ASP up 121%

   TrendForce's latest research estimates memory supply will remain tight through 2027 and has revised up its HBM price outlook, projecting blended ASP to rise about 121% year over year, driven by tight overall supply, a higher share of pricier HBM4 shipments and the gradual ramp of HBM4e in the second half of 2027. The article argues that beyond advanced wafer capacity, stacking yield is the key constraint on effective supply: TrendForce data puts current HBM yield at only 50%-60%, with yields typically falling for each new generation, while JPMorgan data shows 58% of 2025-2028 DRAM capacity additions going to HBM, lifting HBM's share of total DRAM capacity from 19% to 31%. Micron has said it will work with TSMC on mass production of next-generation HBM4E in 2027, and per TrendForce, NVIDIA has been reassessing Rubin Ultra's HBM configuration since Q3 2026, weighing HBM4e 12hi/8hi against HBM4 12hi/8hi.

   虎嗅 AI · 🔥 0
6. 5d
   Samsung offers memory as leverage for AMD foundry orders; AMD stays cautious

   AMD CEO Lisa Su visited South Korea to secure stable high-bandwidth memory (HBM) supply for its AI accelerators and said she would discuss a broader partnership with Samsung executives. Samsung hopes to use its memory supply as leverage to win foundry production of AMD's advanced logic chips; according to Korean outlet THE ELEC, the two firms signed a three-year, non-binding MOU in March 2026 covering HBM4 and advanced DDR5 supply, joint memory optimization for AI and HPC, evaluation of AMD platforms made at Samsung foundry, and advanced packaging cooperation. Su also confirmed AMD has begun shipping next-generation Instinct MI455 chips with 432GB of HBM4 and Helios AI rack systems, with Samsung currently the main HBM4 supplier for the MI400 series, though AMD warned in its latest annual report that shifting production to a new foundry could bring delays, lower yields or weaker performance.

   科技新报 · 🔥 0
7. Oct 7
   Lisa Su in Korea to lock in HBM supply, seeks broader Samsung partnership

   At AMD Gamer Day in Seoul on Oct 7, CEO Lisa Su called South Korea's semiconductor supply chain "indispensable" and said AMD wants three- to five-year (or longer) memory supply arrangements with Samsung Electronics and SK hynix to cope with tight HBM supply, while working with customers to optimize memory usage. She met Samsung's device solutions chief and vice chairman Jun Young-hyun and other executives that day; Taiwan's TechNews reports she also said AMD has begun shipping its next-gen MI455 accelerators and Helios systems using HBM4, backed by Samsung and SK hynix. Reports say Samsung hopes to convert the memory relationship into foundry orders, but Su only spoke of "exploring" memory and foundry cooperation, and neither company disclosed details.

   36氪快讯 · 🔥 0

Experience and discussion from the community

Share my HBM4 experienceAsk about HBM4