힙한 펭미업의 1인칭 시점

Forwarded from Korean Frigate - 해외주식 항해일지

광통신 랠리의 계기가 된 (것으로 보이는) Citi Report
2026.08.05

AI Chipmakers to Pivot from Scale-Up to Scale-Out Amid HBM Shortage

핵심요약
- GPU 당 HBM 용량을 줄이고, 더 많은 GPU를 연결 구성
- Rubin에서 HBM4 12단 → Rubin Ultra에서 HBM4 & HBM4e 8단, 대신 시스템당 GPU 개수는 72개에서 576개로
- GPU당 HBM 용량 줄어들면 KV cache 스토리지를 HBM 외부에서 오프로드 해야 함 → Server DDR5 & eSSD 수요 부각

CITI’S TAKE

Amid persistent memory shortages, we expect AI chipmakers to increasingly scale out by deploying more GPUs with lower HBM content. As the NVLink domain would expand from 72 to 576 GPUs, the aggregate HBM capacity per AI system is projected to increase by 433% to 110.6TB from 20.7TB. We also expect HBM supply to remain tight, as memory suppliers stay disciplined in expanding output amid attractive server DDR5, SOCAMM2 and eSSD economics. We remain constructive on Korea memory suppliers under our coverage.

NVIDIA reportedly considering shifting Rubin Ultra from 12-Hi to 8-Hi HBM — According to SemiAnalysis, NVIDIA is reportedly considering a mainstream Rubin Ultra configuration featuring 8-Hi HBM4 with 192GB of capacity. By comparison, NVIDIA’s official Rubin specification features 12-Hi HBM4 and 288GB of capacity. If adopted, the reported Rubin Ultra configuration would carry 33% less HBM capacity than Rubin and 50% less than the prior market expectation of 384GB for a two-die HBM4E design. Most HBM4 roadmaps have so far focused on higher-capacity 12-Hi and 16-Hi stacks. NVIDIA’s reported consideration of an 8-Hi configuration suggests that it may prioritize reducing HBM costs per GPU and its overall procurement burden over maximizing memory capacity.

Rubin Ultra’s focus may shift to system-level scalability to 576GPUs from 72 GPUs per AI System — The reported specification change suggests that NVIDIA may be shifting away from maximizing memory capacity per GPU and toward expanding the overall GPU domain. NVIDIA’s Rubin Ultra NVL576 connects eight 72-GPU racks into a single 576-GPU NVLink domain. Moving from 12-Hi to 8-Hi HBM could reduce packaging complexity, improve manufacturing yield and throughput, and ease data-center power and cooling constraints, potentially allowing NVIDIA to deploy a greater number of GPUs.

Implication: AI Chipmakers to Pivot from Scale-Up to Scale-Out Amid HBM Shortage — We expect persistent memory shortages, particularly in HBM, to push AI chipmakers away from scaling up individual GPUs and toward scaling out through a greater number of GPUs with lower HBM content. As HBM increasingly becomes a gating component for AI accelerator production, we expect vendors to de-spec mainstream configurations from 12-Hi to 8-Hi as a pragmatic way to stretch limited HBM supply across more GPUs. Despite lower HBM content per GPU, aggregate HBM capacity per AI system would increase by 433% to 110.6TB from the current 20.7TB, as the GPU count per system expands from 72 to 576. Lower HBM capacity per GPU would also require a greater share of KV-cache storage to be offloaded outside HBM, primarily to server DDR5 and eSSDs.

1.48K views12:21