SK hynix presented shared-memory technology for AI servers at OCP Global Summit 2026 and described it in an October 16 newsroom update. The company says its second-generation CMM-DDR5 module offers 256GB capacity using CXL 3.2. It also showed memory-centric rack demonstrations. The development concerns how servers share and reuse data; it is not evidence that every AI workload will become faster or cheaper by a fixed amount.
SK hynix CXL memory: what was disclosed
| Measure | Issuer disclosure |
|---|---|
| Event and announcement | OCP Global Summit, San Jose, October 12–15; issuer update October 16, 2026 |
| Module | Second-generation CMM-DDR5; 256GB; CXL 3.2 |
| Bandwidth | 80GB/s, according to SK hynix |
| Energy efficiency | About 30% improvement versus the previous generation, according to the company |
| NAVER Cloud work | Proof-of-concept validation of memory-sharing use cases |
| MARS | Memory Centric AI Rack System demonstrations with partners |
What pooled memory means
CXL, short for Compute Express Link, is an interface for connecting computing components and memory. In a pooled-memory design, several servers can share a group of memory resources rather than keeping all capacity tied to one server. SK hynix described demonstrations involving reuse of the temporary data, known as the KV cache, that AI language models keep from earlier calculations. Reusing that information can avoid repeating work, but the effect depends on the system and workload.
Company claims are not independent tests
The issuer says the module more than doubles the previous generation’s bandwidth to 80GB/s and improves energy efficiency by about 30%. Bandwidth is a data-transfer measure, not an end-to-end AI response-time result. Better energy efficiency is also not automatically a 30% reduction in a customer’s electricity bill. The comparison is company-reported; CapKet did not reproduce the benchmark or independently validate its test conditions.
The company describes its NAVER Cloud collaboration at proof-of-concept level, meaning a trial of whether a design can work. Its MARS demonstrations combined partner hardware and software around shared memory. Neither a trial nor a booth demonstration establishes availability for all customers. The reviewed announcement does not give this report a verified universal shipping date, price or quantified revenue impact. It is not a forecast of memory sales or a share-price catalyst. The announcement and newsroom index are the same issuer’s materials, not independent corroboration.
Related: TSMC’s reported quarterly revenue and profit, distinct from product-performance claims.
Official sources and verification
Sources observed: 16/10/2026, 11:06:53 IST. Issuer OCP announcement · Issuer newsroom index
Recorded source observation: . This is a source-check timestamp, not a market quote or a new article update.