Bernstein: Maintains Montage Technology (06809) "Outperform" rating, target price HK$520

date
14:23 26/09/2026
avatar
GMT Eight
Bernstein expects Montage Technology's earnings per share to be RMB 3.44 and RMB 5.69 for 2026 and 2027, respectively. The report believes that the target price of RMB 400 reflects the earnings potential from the MRDIMM and CXL product cycles in 2027 to 2028.
Bernstein has released its latest in-depth report on Montage Technology (06809), maintaining an "Outperform" rating, raising its A-share target price to RMB 400, and setting its H-share target price at HK$520. The report argues that as AI applications evolve from single-turn Q&A to agentic models that continuously execute tasks, demand for server CPUs, memory configurations, and high-speed interface chips will rise in tandem, and Montage Technology is poised to benefit from two main themes: MRDIMM upgrades and CXL memory expansion. The core logic behind Bernstein's bullish view is not that AI computing will shift from GPUs to CPUs, but rather that agents bring more work that accompanies model inference. When an agent completes a task, it often needs to retrieve information, call databases, run code, operate browsers, and save session states. Accelerator host CPUs need to handle data feeding and scheduling, sandbox CPUs handle tool execution, and general-purpose server CPUs handle retrieval, databases, and enterprise applications. As these tasks increase, both CPU core counts and memory demand will rise. The report notes that once CPU compute power increases, if memory cannot supply data in time, the efficiency of both the added compute and expensive GPUs may be constrained. For cloud providers, as long as adding CPU and memory configurations can improve the effective output of the entire cluster, there is ongoing incentive to invest. HBM continues to serve high-bandwidth accelerator computing, while local DDR primarily handles data frequently accessed by CPUs, and the two types of demand can grow in parallel. In terms of market sizing, Bernstein adopts a framework of "server CPU shipments DIMMs per CPU value per memory interface chip." The report estimates that excluding Nvidia's own CPUs, global server CPU shipments will grow from approximately 30.6 million units in 2025 to approximately 89.3 million units in 2030; the average number of memory modules per CPU will rise from about 9 to 13.6; and the average value of interface chips per memory module will rise from about US$5.9 to US$16.3. Combining these three factors, the global DDR memory module chipset market is expected to expand from approximately US$1.6 billion in 2025 to approximately US$20 billion in 2030, with a compound annual growth rate of about 65%. Bernstein also cautions that this figure represents the total industry opportunity and does not equate to the revenue Montage Technology can capture; actual realization still depends on customer configurations, product penetration rates, and the company's market share. MRDIMM is seen as the key source of growth elasticity in Montage's core business. Compared with traditional RDIMM, which typically accesses one rank at a time, MRDIMM can use dedicated interface chips such as MRCD and MDB to enable two ranks to supply data in parallel, significantly increasing the memory bandwidth available to the CPU side. According to the report's estimates, the related chip value for ordinary DDR5 RDIMM is about US$6 to US$7, while the related chip value for MRDIMM can reach US$50 to US$70. Bernstein expects that by 2030, MRDIMM will account for about 25% of server DDR memory module shipments, with its MRCD and MDB products contributing a significant proportion of market value. For Montage, this means the company's growth does not depend solely on an increase in server volumes, but will also benefit from rising value per memory interface chip. However, MRDIMM is more suitable for application scenarios with intensive memory requests and bandwidth approaching saturation, and still needs to pass CPU, motherboard, firmware, thermal, and complete system certification, so its penetration pace still needs to be verified through actual customer deployment. In addition to MRDIMM, CXL memory expansion is another new opportunity Bernstein favors. Montage already has related products supporting CXL 2.0, PCIe 5.0, and DDR4 and DDR5. CXL can help next-generation servers reuse some DDR4 memory to expand capacity at lower cost, making it suitable for storing waiting agent environments, infrequently accessed vector data, and cache content. The report estimates that under specific price assumptions, the cost of native 128GB DDR5 is about US$3,500, while a solution that reuses DDR4 with a controller and expansion card costs about US$920. Bernstein expects the market for CXL controllers related to DDR4 reuse to grow from about US$180 million in 2027 to about US$8.2 billion in 2030, with the total CXL chip market size potentially approaching US$9.9 billion. The competitive landscape is also an important support for Montage's high valuation. Data cited in the report shows that Montage, Renesas, and Rambus together account for more than 90% of the global core memory interface chip market, with Montage holding about a 37% share. The industry typically requires 18 to 24 months from chip design to customer certification and mass production introduction, and long-term validation and stable supply capabilities constitute high barriers. On valuation, Bernstein expects Montage Technology's earnings per share in 2026 and 2027 to be RMB 3.44 and RMB 5.69, respectively. The report argues that the RMB 400 target price reflects the earnings potential of the MRDIMM and CXL product cycles from 2027 to 2028. Going forward, MRDIMM customer certification and mass shipments, CXL's transition from trial use to procurement, gross margin performance, and changes in technology roadmaps such as DDR6 and LPDDR will become key indicators for validating the company's growth logic.