Leak: OpenAI Asked Samsung for 4-Hi HBM Config but Was Rejected Amid Supply Constraints
zephyr_z9 · x · 2026-09-17
Leaker zephyrz9 claims the 4-Hi HBM stack configuration faces base-die supply constraints and resistance from memory makers, even though AI labs strongly favor it. According to the post, OpenAI asked Samsung for this config but was rejected for now.
He also does the math: holding one copy of model weights at fp4 requires 20TB of HBM. Using Huawei's hypothetical 4-Hi XPU (4GB × 4-Hi × 6 stacks = 96GB per accelerator), a 4096-accelerator scale-up node would carry over 384TB of HBM.
Related event: Rumor: Huawei to expand supernode to 4096 cards with 384TB HBM(2 posts)→
More from Infra
- Nebius GPU rate hikes and Fed hikes are connected: AI debt is pushing up yields — kevinsxu · 2026-09-17
- Weaviate 1.39 adds 4-bit Rotational Quantization, cutting heap memory 45% — CShorten30 · 2026-09-17
- Single AMD R9700 Hits 153 tok/s Running Qwen3.8 27B NVFP4 After Optimization — whodoneit1 · 2026-09-17
- Unsloth Desktop Launches: Train & Run 500+ Local Models, 70% Less VRAM — danielhanchen · 2026-09-17
- Back-of-envelope: 40T-param sparse models are inevitable, DeepSeek on track — teortaxesTex · 2026-09-17
- He ditched Codespaces for exedev: 50 VMs, no per-VM charge, agents on tap — swaroopch · 2026-09-17