The take
Samsung and Micron have both advanced to volume shipping of HBM4 for Vera Rubin, while SK Hynix holds an estimated 70% HBM4 bit-share — consolidating the supply picture around three qualified vendors. Simultaneously, a severe DRAM shortage has pushed 8GB module costs up 8.5x, squeezing smaller manufacturers out of the market. South Korea's $576B investment commitment with Samsung and SK Hynix signals capacity relief, but execution risk is already pressuring both stocks.
What moved this week
TSMC CoWoS CoWoS CoWoS capacity — up — Q4 2026 ~105k–127k wafers per month by Q4/end-2026 (JPMorgan at 105k vs. a separate estimate at ~127k) vs Q4 2027 141k wafers per month by 4Q27 (
source)
SK Hynix HBM bit share — down — ~55–58% (readings range 50–62%) vs 2025 55% (
source)
TSMC CoWoS CoWoS pct locked — flat — Fully booked; top 3 customers account for >85% of capacity vs 2026-06 fully booked through mid-2026 (
source)
SK Hynix HBM4 HBM bit share — up — 2026 ~70% vs ~62% (
source)
SK Hynix HBM HBM bit share — up — 2025 55% vs 70% (
source)
SK Hynix HBM3E hbm spot price — down — 2026 ~$10/GB (
source)
Supply-chain & pricing signals
Huawei overtakes Nvidia in China AI chip market share — U.S. export controls are accelerating domestic Chinese chip adoption, fracturing the global AI chip addressable market. (
AP News)
South Korea announces $576B AI chip investment with Samsung and SK Hynix — New fabrication sites and a packaging cluster near Seoul target HBM and memory capacity expansion within a 2–3 year horizon. (
CNN)
Severe DRAM shortage drives 8.5x cost surge on 8GB modules — Only large-volume buyers can secure adequate supply; smaller manufacturers face margin compression or spec reductions. (
CNBC)
China expands export controls on Japanese defense-linked entities and dual-use semiconductors — Japan's defense and aerospace sectors face new procurement risks, accelerating regional chip sourcing fragmentation. (
CNBC)
Soitec and ZenSemi scale 300mm BCD-on-SOI for AI datacenter power electronics — Partnership targets supply-chain diversification for power management chips in power-intensive AI infrastructure. (
Manila Times)
HBM watch
Samsung × NVIDIA Vera Rubin (HBM4 12-Hi): Qualified
→ Volume shipping — Samsung announced world-first HBM4 mass production shipment at 11.7 Gbps with 12-layer stacking. (
source ·
timeline)
Micron × NVIDIA Vera Rubin (HBM4): Qualified
→ Volume shipping — NVIDIA CEO confirmed Micron passed HBM4 certification for Vera Rubin as of June 2026. (
source ·
timeline)
Micron × NVIDIA Vera Rubin (HBM4 16-Hi): Not started
→ Sampling — Micron shipped 48GB 16-Hi HBM4 samples to customers, demonstrating 16-die stacking capability. (
source ·
timeline)
Samsung × AMD (HBM4): Not started
→ Qualified — Samsung passed HBM4 qualification for AMD per chip industry sources cited by Reuters. (
source ·
timeline)
SK Hynix × NVIDIA Vera Rubin (HBM4 16-Hi): Not started — SK Hynix unveiled a 16-layer 48GB HBM4 device at CES 2026 with mass production planned for Q3 2026. (
source ·
timeline)
This week's analysis
2024 Capacity Expansion: AI Accelerators, GPUs, HBM, and Wafer Starts Per Month at TSMC, Samsung, and Intel — A comprehensive analysis of 2024 semiconductor capacity expansion for AI accelerators, GPUs, and HBM — covering wafer starts per month, advanced packaging constraints, and strategic moves by TSMC, Samsung, and Intel.
AI engines are citing
Tool spotlight
Cost Bridge — Side-by-side manufacturing-cost breakdown across 13+ AI chips.
Sources this week
Sourced public intelligence, not insider data; status changes are human-reviewed before publication. Full timelines on the HBM Qualification Tracker.