SK Hynix confirms HBM4 volume ramp for September — 432GB modules going to NVIDIA, AMD, and three hyperscaler-custom buyers
SK Hynix confirmed its HBM4 volume production ramp for September 2026 this week, with 432GB stacks shipping to NVIDIA, AMD, and three undisclosed hyperscaler-custom accelerator buyers. The September ramp materially shortens the bottleneck window for HBM-constrained accelerator launches, but spot pricing remains 38% above the equivalent-density HBM3e tier through Q4.
The September ramp date is the more material number than the per-stack capacity. AMD's MI400 launch in H2 was paced against HBM4 availability; NVIDIA's next-generation Rubin parts are similarly HBM4-gated. SK Hynix shipping at volume in September means the second half of 2026 frontier-accelerator ramp is no longer memory-bound for the first time since the 2023-2024 HBM shortage.
The three hyperscaler-custom accelerator buyers are the more interesting story. Google TPU v6, AWS Trainium 3, and Meta's MTIA v3 are the candidate consumers, and each parts list now reads HBM4-432GB. That means the in-house silicon programs at the three frontier-AI hyperscalers are entering the same memory-capacity tier as the merchant accelerator parts.
SK Hynix — HBM4 volume ramp announcement → · DRAMeXchange — HBM4 supply timeline 2026 → · SemiAnalysis — What HBM4 ramp does to accelerator margins →