Samsung Secures Half of Nvidia’s SOCAMM 2 Orders for 2026

Samsung has recently entered Nvidia’s HBM supply chain with its HBM3E products and is now in price negotiations for HBM4. Amid this progress, the company has achieved another major breakthrough: securing half of Nvidia’s SOCAMM 2 production for next year. This shows the firm’s solid position in the rapidly growing AI memory market.

Samsung wins SOCAMM 2 deal as Nvidia wants to power Vera-powered systems

According to a report from Hankyung, Nvidia is asking around 20 billion gigabytes of SOCAMM 2 from the global DRAM industry. Samsung receives orders for about half that amount, roughly 10 billion gigabytes. Industry estimates that the company will need to make around 30,000 to 40,000 wafers per month to meet supply demand. This is only about 5% of its total DRAM capacity. Meanwhile, SK Hynix may handle most of the remaining volume (around 60 to 70%), with Micron receiving a smaller share.

Samsung has recently improved its 10nm-class fifth-generation (1b) DRAM, which is used to make SOCAMM 2 modules. Earlier this year, the company lagged behind its rivals in 1b and 1c process performance, but rapid design upgrades and stable yields now helped it beat Micron.

Nvidia initially planned to use SOCAMM with the Grace CPU in its GB300 server platform. However, the first version did not go well due to heat and performance issues. The new SOCAMM 2 module will now debut with Nvidia’s new Vera CPU next year. It offers 50% higher capacity and 20% improved power efficiency than the first-generation SOCAMM module.

For the uninitiated, SOCAMM, short for Small Outline Compression Attached Memory Module, is a new type of DRAM module that powers AI data centers. Also known as the second high-bandwidth memory (HBM), it sits next to the CPU and coordinates tasks for the GPU. Nvidia first debuted SOCAMM at its GTC 2025 conference in March, where Samsung, SK Hynix, and Micron Technology showed their early prototypes.