Samsung Q2 profit jumps ~18-fold on AI memory; Unpacked debuts on-device AI

Samsung's preliminary Q2 2026 operating profit surged roughly 18-fold year-over-year to about 86 trillion won ($56.35 billion), driven by soaring demand for AI memory — the HBM and high-density DRAM that feed data-center GPU clusters. The scale of the jump underscores how the AI build-out has become the dominant profit engine for memory makers, echoing NVIDIA's role on the compute side.

At its Unpacked event on July 10, Samsung leaned into the device side, introducing on-device AI powered by custom silicon (reportedly a 3nm chip running 7B-class models) and a revamped software stack that executes advanced generative models locally on Galaxy devices — no cloud round-trip. The pitch pairs privacy and latency benefits with 'zero-touch' toolkits for Python, Kotlin, and Java.

The on-device angle drew genuine enterprise interest: developers on Perplexity/HN threads flagged local execution as a privacy and compliance win, though consumer enthusiasm was more muted — most users don't distinguish on-device from cloud AI. It positions Samsung against Apple Intelligence's on-device story and Qualcomm's edge-AI push.

The two threads connect: Samsung profits from selling memory into the cloud AI boom while simultaneously betting that inference migrates to the edge, where its silicon and device install base give it leverage. The risk is memory cyclicality — an 18-fold swing cuts both ways — and whether on-device models stay good enough as frontier cloud models keep improving. Watch HBM pricing and whether Galaxy's local AI meaningfully changes usage.