HBM demand is almost entirely driven by AI accelerator chip attach rates. Each NVIDIA H100 GPU uses 80 GB of HBM3E; the Blackwell architecture uses significantly more. Growth will be front-loaded as AI infrastructure buildout peaks, then moderate as the install base matures. The bull case assumes continued AI infrastructure buildout at current pace with HBM4 adoption accelerating across NVIDIA, AMD, and hyperscaler custom silicon. The bear case assumes AI capex moderates in 2028–2029 as hyperscalers face utilization pressure or alternative memory architectures (e.g., CXL-attached DRAM) reduce HBM dependency per accelerator.
“HBM supply remains tight — demand exceeding available capacity through FY26. All HBM3E production sold out; HBM4 qualification underway at key AI chip customers.”
“HBM3E demand from NVIDIA H100/H200 and next-gen Blackwell GPUs accelerating. SK Hynix maintaining HBM market leadership with ~50% share; capacity fully committed through FY26.”
“HBM4 roadmap on track. Custom logic integration advancing for major hyperscaler customers. Pricing premium over standard DRAM remains elevated at 4-5x.”