The three companies that collectively dominate the global memory chip market have decided they don’t want to build a key piece of the puzzle themselves. Samsung Electronics, Micron Technology, and SK hynix are all scaling back or outright abandoning their in-house efforts to develop CXL controllers, the specialized silicon that manages next-generation memory expansion in data centers.

Who’s doing what, exactly

Micron has gone the furthest, fully halting its internal R&D and adopting a CXL 3.0 controller solution from Primemas, a fabless semiconductor company that has been sampling its controller technology since mid-2025.

SK hynix completely terminated its CXL controller project and redirected those engineering resources toward processing-in-memory, or PIM, research. Rather than building the traffic cop that directs data between processors and memory, SK hynix wants to put computing power directly inside the memory chips themselves.

Samsung is keeping its controller project alive but only for internal experimentation. For anything commercial, the company plans to source controllers through outside partnerships.

The CXL backstory

CXL, or Compute Express Link, is an open standard that lets processors share memory pools across a data center more efficiently. It’s supposed to be transformative for AI workloads, cloud computing, and any application that needs massive, flexible memory capacity without the astronomical cost of simply cramming more DRAM onto every server.

The three memory giants had collectively committed over $12 billion to CXL-specific development through 2027. But the anticipated ramp-up for CXL 3.x market volume has slipped, with meaningful adoption now projected for 2027 and 2028, pushed back by production delays that have plagued the ecosystem.

Primemas has emerged as a key beneficiary of this recalculation. The fabless firm’s collaboration with Micron on CXL 3.0 controllers represents a significant validation of the outsourcing model.

Why this matters beyond semiconductors

Data center infrastructure is the physical backbone of everything from AI training clusters to blockchain node operations and crypto exchange matching engines. CXL technology, when it actually arrives at scale, could meaningfully reduce the cost of running large-scale distributed computing operations.

With CXL 3.x volume now expected to ramp primarily in 2027 and 2028, companies that were pricing in earlier adoption may need to recalibrate. The $12 billion already committed across the three majors isn’t going away, but the return on that investment is arriving later than planned.

The PIM pivot at SK hynix deserves particular attention. Processing-in-memory represents a fundamentally different architectural approach that could eventually compete with or complement CXL rather than simply feeding into it.