Samsung's Memory Comeback: HBM4 Chips Head to Nvidia in February

Samsung Secures Major AI Memory Win

Samsung Electronics is set to ship its next-generation HBM4 memory chips to Nvidia starting in February 2026, marking a potential turning point for the South Korean tech giant after struggling to compete with rival SK Hynix in the high-bandwidth memory market.

The chips have passed final quality testing from both Nvidia and AMD, achieving speeds of eleven point seven gigabits per second in internal testing, surpassing Nvidia's requirements for its upcoming Rubin AI accelerator platform.

Technical Innovation Drives Performance

Samsung's HBM4 uses a more advanced four nanometre process for its logic die, compared to SK Hynix's twelve nanometre process from TSMC. The company claims this provides a forty percent improvement in energy efficiency over competitors.

The Rubin GPU platform, unveiled at CES 2026, will feature eight stacks of HBM4 memory delivering two hundred and eighty-eight gigabytes of capacity and twenty-two terabytes per second of bandwidth, nearly tripling memory bandwidth compared to the previous Blackwell generation.

Market Share Recovery Predicted

Samsung's HBM market share is forecast to surge to thirty-five percent in 2026, more than doubling from an estimated sixteen percent in 2025. SK Hynix remains the dominant supplier with over fifty percent market share, though Samsung's co-CEO Jun Young-hyun noted that customers have remarked that Samsung is back concerning HBM4.

Record Profits Signal Momentum

The HBM progress comes amid Samsung's broader turnaround driven by surging AI demand. The company projected a record quarterly operating profit of twenty trillion won for Q4 2025, the first time a South Korean company has exceeded this threshold. The memory chip division accounted for more than seventy percent of total operating profit.

The Race Continues

Both Samsung and SK Hynix plan to begin HBM4 mass production in February 2026, with full-scale volume supply expected around June. Nvidia has already requested sixteen-layer HBM memory chips by Q4 2026, presenting formidable technical challenges as wafer thickness must be reduced from fifty to around thirty micrometres.