三星、SK海力士、美光据称停止自研CXL控制器并转向外部Fabless方案

Memory Big Three Scrap In-House CXL Controller Chip Development

The world’s three largest memory chipmakers have all halted in-house development of controllers for Compute Express Link (CXL), a next-generation semiconductor technology in which they had invested heavily. The decision reflects concerns over a cannibalization dilemma: aggressively pushing proprietary chips into the market could undermine demand for commodity DRAM modules, or DIMMs, their core source of revenue. The memory giants are rapidly shifting to an outsourcing model, adopting chips designed by specialized fabless companies instead.

According to semiconductor industry sources on the 20th, Samsung Electronics, SK hynix, and Micron have scaled back or scrapped plans to commercialize controllers for CXL expansion devices. Fabless companies such as Montage Technology, Astera Labs, and PrimeMass are moving in to fill the gap.

Micron and SK Throw in the Towel, While Samsung Tests the Waters

Micron was the first to abandon in-house development. The U.S. chipmaker shut down its proprietary controller R&D operation and adopted a solution from PrimeMass. PrimeMass solutions now also appear in Micron’s product catalog.

SK hynix has likewise formally notified key partners that it is discontinuing its in-house CXL controller development project. The company is reassigning relevant personnel to processing-in-memory (PIM), a next-generation computational memory technology. The move reflects its intention to concentrate limited R&D resources on a more certain source of future growth.

Samsung Electronics is using its internally developed CXL controller only for research within the development team. The team is focusing on unproven areas, including CXL solutions based on low-power DRAM, or LPDDR. For products sold externally, Samsung is reportedly purchasing controllers from fabless suppliers.

This marks a departure from Samsung’s original plan. The company had initially intended to launch modules combining its proprietary controller with a CXL Memory Module, or CMM, but has since removed the formal commercialization of its in-house controller from its official roadmap.

“Samsung’s internal product-planning division has removed the in-house controller from its formal commercialization projects and retained it solely as an advanced R&D initiative,” said a semiconductor industry source familiar with the matter. “The development team is now concentrating on areas with uncertain near-term commercial viability, such as running mobile DRAM, or LPDDR, over CXL. In effect, Samsung has returned to a market-exploration phase.”

“Promoting CXL Could Destabilize DRAM”—The Cannibalization Dilemma

According to industry sources, all three memory manufacturers originally envisioned selling integrated CXL modules similar to Samsung’s proposed design. Their plan was to combine proprietary controller chips and DRAM on a single board and sell the resulting system as a high-priced finished product.

Major data-center customers, however, wanted something different. To reduce the enormous cost of building infrastructure, they preferred a disaggregated architecture: installing the CXL controller separately on the system motherboard and inserting widely available, inexpensive commodity DRAM modules into slots behind it for reuse.

Analysts say this is where the business models came into conflict and why the manufacturers changed course. If memory companies insisted on expensive integrated products built around proprietary controllers that had required substantial development spending, cost-conscious customers would be reluctant to buy them. Worse still, such products could reduce demand for the commodity DRAM modules that the companies already sell in large volumes. In other words, aggressively promoting finished CXL products could cannibalize the DIMM market that serves as their largest source of revenue.

“If a memory manufacturer’s finished CXL expansion product must compete directly with its own core cash cow—DIMMs—it becomes difficult to pursue the business aggressively,” said a semiconductor industry official. “Management concluded that there was little practical benefit in forcing the commercialization of in-house controllers while risking a conflict with the company’s existing core business.”

Experts stress that the three companies’ decisions should not be interpreted as a retreat from or abandonment of CXL technology and the CXL market itself. Rather, they see the move as a way to reduce risk during the market’s uncertain early stages and return to the semiconductor industry’s naturally efficient division of labor.

“Instead of trying to monopolize control of CXL, the memory manufacturers have opted for a pragmatic division of labor with fabless companies,” another industry official said. “As the global semiconductor market becomes more sophisticated, this optimized ecosystem—in which specialized fabless companies handle design and memory manufacturers focus on production—is expected to accelerate further.”