Macro

Multi-year optical demand and pricing power validated through H2 2027 across major hyperscaler customers.

Neocloud datacenter buildout requires significant power supply infrastructure investment and buildout.

Inference workloads projected to grow from 50% to 65% of total compute allocation next year, reflecting accelerating shift toward inference-focused architectures.

Intel's analysis suggests CXL connections are necessary for effective KV cache offloading beyond GPU memory, positioning CXL as a critical infrastructure layer for accelerator efficiency.

Optical components (CPO/LPO) operating at capacity limits with strong pricing power, indicating tight supply-demand balance.

SK Hynix is investing tens of trillions of won in a new memory fab in Japan, signaling commitment to expanding HBM/DRAM capacity amid AI infrastructure demand.