Samsung Electronics, SK hynix, and Micron have scaled back in-house development and commercialization of controllers used in CXL memory expansion devices, ZDNet Korea reported on July 20, 2026. What the three companies are said to be reconsidering is the in-house approach to control chips for commercial products—not CXL memory products themselves. CXL-compatible products remain on each company's current product pages and at 2026 exhibitions. The more accurate way to characterize the shift is a move away from an integrated approach built around self-designed controllers, toward a division of labor that incorporates designs from specialized companies.
Three Companies, Three Different Approaches to Scaling Back In-House Development
According to ZDNet Korea's report, Micron has streamlined its proprietary controller R&D line and adopted products from Primemas. SK hynix reportedly informed key partners that it is discontinuing in-house development and is reassigning personnel to the PIM (Processing-in-Memory) field. Samsung is said to be limiting the use of its in-house controllers to preliminary research within its development division, with a policy of sourcing from fabless companies for products sold externally.
However, none of the three companies have officially announced changes to their in-house development plans. Samsung has not completely halted its research either, and reportedly continues to validate configurations such as connecting LPDDR to CXL. Describing all three companies collectively as having "abandoned CXL development" would go beyond the facts as reported.
The reason cited in the report is competition with existing server DIMMs. Even if a company pushes integrated modules with expensive proprietary controllers built in, sales won't grow if customers prefer configurations that use standard DIMMs. Moreover, there are reportedly concerns that CXL products could cannibalize a company's own DIMM demand. Notably, ZDNet Korea did not provide specific figures for development costs or lost revenue, so there is no confirmed data backing up any claim of "billions of dollars" in losses.
Separated Designs That Leverage Standard DIMMs, Rather Than Integrated Ones
CXL memory requires a controller that translates CXL commands arriving from the CPU into DRAM read/write operations. The CXL Consortium defines Type 3 devices—which provide memory under CXL 2.0—as memory controllers, and states that Type 3 also handles memory pooling, which allows multiple hosts to share capacity.
In the integrated approach, the controller and DRAM are combined into a single module, such as an E3.S form factor. This is easy for adopters to treat as a finished product, but the available capacity and DRAM options are constrained by the vendor's product design. In the separated approach, by contrast, the controller sits on the motherboard or an add-in card, with standard RDIMMs inserted into slots downstream. This allows DRAM to be sourced from existing supply chains, and capacity can be reconfigured more easily afterward.
Astera Labs already lists a CXL 1.1/2.0-compatible expansion card in its product lineup, featuring four DDR5-5600 RDIMM slots and up to 2TB of capacity. Even as standalone controllers, it offers multiple models that connect DDR5 over two channels and handle up to 2TB. The separated approach is not merely a concept on paper. That said, total server-level costs need to be compared by including software support and workload-specific performance, in addition to the controller and board itself. The mere ability to use standard DIMMs does not necessarily guarantee lower costs.
CXL Products Remain: From 128GB to 256GB
Samsung's current product, the MD220, houses 128GB and 256GB of DDR5 in an E3.S 2T form factor, connected via CXL 2.0 over PCIe 5.0. Test tools and white papers covering CXL memory have been added to the company's product page, dated between March and June 2026. Even as the source of controllers changes, development of CXL memory and surrounding software continues.
Micron also currently lists the CZ122 in its AI data center product lineup. It offers 256GB of capacity, with configurations expandable up to 2TB and 36GB/s of bandwidth. Furthermore, in June 2025, Primemas announced plans—through Micron's CXL ASIC Validation Lab program—to pair a CXL 3.0-compatible Hublet controller with 128GB RDIMMs. It has already been publicly disclosed that Micron is evaluating external designs and working out compatibility and reliability with its own DRAM.
SK hynix announced in April 2025 that it had completed customer certification for its 96GB CXL 2.0-compatible CMM-DDR5, with a 128GB version currently undergoing certification. At COMPUTEX in June 2026, the company also exhibited a CXL 2.0+-compatible CMM-DDR5. What can be gleaned from official information is not a picture of the three companies abandoning CXL, but rather one where they continue selling DRAM and finished modules while reconsidering how control chips are made.
The Spotlight Shifts from Major Memory Makers to Specialized Controller Companies
If the scaling back of in-house development becomes entrenched, some of the value will shift to companies like Primemas and Astera Labs. Major memory makers will focus on DRAM manufacturing, binning, and module validation, while specialized companies will handle keeping pace with CXL standards and supporting multiple DRAM configurations. Since customers can choose memory and controllers separately, they become less tied to a single manufacturer's product roadmap.
At the same time, this division of labor alone will not determine the timeline for CXL adoption. Samsung itself has explained that standards newer than CXL 2.0 have not yet been sufficiently commercialized, and that the overall market—including peripheral devices that connect CPUs and memory—is still maturing. Even if standard-compliant chips exist, the timing of large-scale adoption remains uncertain until the market, including peripheral devices, matures.
The figures worth verifying from the second half of 2026 onward are the number of mass-produced products adopting externally made controllers, and how many units have moved from customer certification to actual deployment. If those numbers grow, it will confirm that the division of labor incorporating externally made controllers and the expansion of mass-produced CXL products are advancing simultaneously.