SK Hynix has started pumping out 192GB SOCAMM2 modules to feed the hungry beasts, which are Nvidia’s next AI data centres.
At CES 2026, SK hynix said it had delivered next-gen memory, including SOCAMM2, to Nvidia for upcoming AI data centre platforms. Four months later, it is claiming mass production is live and ready for Nvidia’s next wave.
This is being lined up for Nvidia’s Vera Rubin platform, with SK hynix joined by Samsung and Micron, so the supply chain does not fall over when demand spikes in the agentic AI era.
SK hynix says the new 192GB SOCAMM2 is built on its 1cnm process, which it describes as the sixth generation of its 10-nanometre-class technology, and uses LPDDR5X low-power DRAM.
The firm claims the module delivers more than double the bandwidth and more than 75 per cent better power efficiency than conventional RDIMM, which is a polite way of calling the old stuff a power hog.
SK hynix president and head of AI Infra (CMO, chief marketing officer) Justin Kim said: “By supplying the 192GB SOCAMM2, SK hynix has established a new standard for AI memory performance. We will solidify our position as the most trusted AI memory solution provider through close collaboration with our global AI customers.”
The company reckons this will ease memory bottlenecks during training and inference for large language models with hundreds of billions of parameters, where bandwidth and power limits love to ruin everyone’s day.
It is pitching SOCAMM2 as the sort of low-power server memory that matters more as the market leans harder into training workloads, while cloud service providers keep demanding faster kit without the electric bill turning into a horror story.
SK hynix says it has been widening its supply portfolio and stabilising mass production early, which reads like a pre-emptive apology for what happens when everyone tries to buy the same modules at once.
SOCAMM2 is essentially low-power memory, previously used in phones, repackaged for servers as a primary memory option for next-generation AI boxes.
The name is Small Outline Compression Attached Memory Module 2, an AI server-optimised module based on LPDDR with a slim form factor and a compression connector designed to boost signal integrity and make swaps less painful.
There is even a nod to “Perf/Watt ‘Optimised’ AI rack scale solutions” in the wider ecosystem, with SOCAMM2 turning up as the memory idea everyone wants to claim first.
For a bit of extra weirdness, the press images show Solidigm SSDs wearing ‘SOCAMM2’ labels.
