Offshore

Jukan RT @MarkosAAIG: Good post by @jukan05 . Not much to add here. Good to see that Micron also has its share in HBM4. Strengthens NVIDIA’s multi-sourcing strategy to the fullest. these days there are a lot of headlines being thrown around and we need to…

stomers (wafer-level delivery vs. individual chip-level delivery), the impact varies, but generally, lower yields mean lower profitability.

Production capacity is also still insufficient. The production capacity for 1c DRAM — the fundamental building block of HBM4 — is at approximately 70,000 wafers per month, which is about 10% of Samsung's total DRAM capacity. Although the company has recently begun expanding its Pyeongtaek Plant 4, production capacity won't increase to around 190,000 wafers until a year from now. Manufacturing HBM4 also requires "cutting-edge packaging" that makes the GPU and HBM function as a single chip. This process further reduces the number of surviving chips. From Samsung's perspective, even if it wanted to produce more HBM4, it simply cannot at this point.

Commodity DRAM Prices Have Risen to HBM3E Levels… Even Higher Profitability

Demand remains strong for HBM3E 12-layer products, which will serve as the market's mainstream through the first half of this year. Samsung passed Nvidia's HBM3E 12-layer qual test around September–October of last year, but actual shipment volumes are reportedly not large. Instead, Samsung is reportedly receiving a flood of orders for "HBM3E 12-layer" from Broadcom, which co-designs Google's AI accelerator TPU, among others. For Samsung, there is little incentive to cut production capacity for HBM3E 12-layer — which uses 10nm 4th-generation (1a) DRAM — and go all-in on HBM4.

Prices for high-margin commodity DRAM used in servers, mobile devices, and PCs are surging. According to global investment bank Goldman Sachs, commodity DRAM prices this year are around $1.25 per gigabit (Gb), or about $10 per gigabyte (GB; 1 GB = 8 Gb). This is not far from HBM3E (5th-generation HBM) 12-layer products, the current mainstream in the HBM market. Since commodity DRAM does not require the cutting-edge packaging that HBM does, its profitability (margin) is reportedly incomparably higher than HBM.

For Samsung, which holds a production capacity advantage of over 1.2x compared to competitors in DRAM, a strategy of proving its technological recovery with the title of "first-ever HBM4 shipment to Nvidia" while focusing on the more profitable commodity DRAM business was deemed the rational choice.

Nvidia Also Needs SK hynix

From Nvidia's perspective, Samsung's strong showing is welcome news, but the company may have judged that allocating a larger share to SK hynix — a long-time partner it has worked closely with — is the more "stable" option. SK hynix is reportedly making progress in the qual testing process. For Nvidia, which has committed to deploying large volumes of HBM4 in its new AI accelerator Vera Rubin in the second half of the year, the stable supply from SK hynix — which has the largest HBM production capacity — is of paramount importance.

That said, future variables do exist. If a particular supplier's HBM4 fails to deliver adequate performance, Nvidia simply cannot accept the products. There is precedent: last year, when Samsung repeatedly failed Nvidia's HBM3E 12-layer qual test, the shares allocated to SK hynix and Micron increased. More recently, reports have emerged that Micron is experiencing difficulties with its HBM4 shipments to Nvidia, leading to projections that Samsung's share could rise to as high as 30%. - Jukan tweet

1 view02:32