Samsung forecasts $80.1B Q3 operating profit, up 782% on AI memory demand

The profit surge reflects the high-bandwidth memory boom. Every AI accelerator, from NVIDIA's GPUs to Microsoft's Maia, needs stacked HBM, and supply remains tight. Samsung, which lagged SK Hynix in early HBM generations, appears to have caught up in volume and pricing.

At OCP, Samsung laid out its roadmap. HBM5 promises higher bandwidth. Custom HBM integrates customer-specific logic in the base die. Processing-in-memory runs some compute inside memory to cut data movement. CXL memory pooling with AMD lets servers share memory capacity, easing the memory wall for large-context inference.

On the device side, Samsung showcased Gauss-powered AI across TVs, appliances and the Galaxy Z Fold8 at KES 2026. Community discussion notes that Samsung's group-wide adoption of Gemini, ChatGPT and Claude marks a shift away from relying on in-house Gauss.

Caveats: preliminary guidance, and memory is cyclical. Bank efforts to offload AI data-center debt hint at bubble risk. Watch the final results and HBM5 qualification with NVIDIA.