SK hynix used Dell Technologies World 2026 to put its entire AI memory and storage line in one booth, from 16-layer HBM4 down to a client SSD it has been supplying to Dell since April.
The HBM display carried actual units rather than renders, specifically 48GB 16-layer HBM4, 32GB 12-layer HBM4 and 36GB 12-layer HBM3E, set beside GPU system models for AI data center servers. Showing a 16-layer HBM4 part inside a server customer's event is a different signal from showing one at a memory conference. It puts the stack in front of the people writing bills of materials.
Server DRAM got the larger stand. The company showed 192GB SOCAMM2, a 256GB 3DS RDIMM it calls the industry's largest capacity for AI workloads, a 64GB RDIMM built on what it calls the world's first 10-nanometer sixth-generation process, a 96GB RDIMM aimed at memory-bound server work, and a CMM-DDR5 module running on CXL. Both superlatives are the company's own, and this release attaches no comparison figure to either. That is not how SK hynix always writes these: its MWC 2026 showcase gave HBM4 as 2.54 times the bandwidth of the previous generation with power efficiency better by more than 40%.
Storage is where something genuinely new sat. SK hynix describes its PEB210 E1.S 9.5mm as "an industry first to support Direct Liquid Cooling (DLC)", which matters more than the phrasing suggests. E1.S is the thin ruler format that took over dense flash servers, and it has stayed an air-cooled island inside racks that are otherwise moving to liquid. A drive that accepts a cold plate removes one of the remaining reasons to keep fans in a liquid-cooled chassis. Around it sat the PS1110 E3.S for high-end NVMe work, plus PS1010, PS1101 and PEB110 parts across E3.S and E1.S.
The most telling exhibit was not a component. Between the two stands the company placed a Dell R7625 server with its own 64GB DDR5 RDIMM and PS1010 E3.S installed inside it. A populated host system is a different claim from a parts display, and it is the one exhibit here that implies a design win rather than a qualification sample.
Two NAND product planning technical leaders, Cheolhoon Yang and Hyunsoo Kang, took a breakout session on the first day titled "Enterprise and Client SSD Industry Trends and Portfolio for Dell." Yang flagged the "evolution of form factors" and "advancement of cooling technologies" as key technical directions for memory efficiency in AI data centers. Decoded, that is a NAND planner telling a server audience that the thermal envelope rather than the interface now sets how much flash fits in a chassis.
The client side leaned on the PQC21 M.2 2230, the 321-layer QLC part with SLC caching that has been going into Dell systems since April. Putting it in an enterprise booth is a sales decision. The company wants the AI PC socket and the server socket sold as one account rather than two.
Bootnote: the 1c label on that 64GB RDIMM is the sixth step in a naming scheme that has run out of letters twice. SK hynix's own footnote spells the sequence out, from 1x to 1y to 1z, then a restart at 1a, 1b and now 1c. Six generations have shipped under the same 10-nanometer class heading without the number in it ever moving.





Leave a Reply