Samsung Electronics and SK Hynix are moving forward with CXL (Compute Express Link) memory production, even as major AI researchers cast doubt on the technology's ability to replace high-bandwidth memory (HBM) in practical AI applications. The two South Korean memory giants are positioning CXL as a complementary layer in AI data center memory hierarchies, not as a direct HBM substitute, according to recent industry discussions.

Why Are AI Leaders Skeptical About CXL Replacing HBM?

At the AI Infrastructure Summit in California, researchers from OpenAI and Intel raised significant concerns about CXL's practical utility for AI model execution. Daniel Morris, an OpenAI researcher focused on AI accelerator design, stated that he has yet to identify a compelling use case for CXL in actual AI model workloads, noting that CXL's role may be largely limited to storing inactive data that is rarely accessed by large AI models.

Vidhya Thyagarajan, Intel's head of AI System-on-Chip architecture, pointed to a critical technical limitation: the data transfer speed between GPUs via CXL is "by no means as fast as HBM." This bandwidth gap makes CXL better positioned as a complement to secondary storage than as a substitute for the high-speed memory that AI accelerators require.

How Are Samsung and SK Hynix Positioning CXL in the Memory Stack?

Rather than viewing CXL as a replacement technology, both companies are integrating it as a new tier within the broader AI memory hierarchy. Samsung plans to begin mass production of its CXL Memory Module (CMM-D) 3.0, based on the latest CXL 3.2 specification, by the end of 2026. Meanwhile, SK Hynix has already debuted samples of its second-generation CMM-DDR5 based on CXL 3.2 and has begun supply talks with potential customers, though it has not yet disclosed a firm timeline for mass production.

SK Hynix's approach demonstrates how CXL can enhance AI inference efficiency. In June, the company unveiled its "Inference Tier Memory Expansion" (ITME) architecture, which inserts CXL hybrid memory between DDR and SSD in the traditional HBM-based AI memory stack. By expanding the conventional HBM, DDR, and SSD hierarchy into HBM, DDR, CXL memory, and SSD, SK Hynix improved AI inference efficiency by 35.7%. Under this new architecture, CXL memory serves as a memory hub that predicts hot data needed by HBM and DDR, retrieving that data from SSDs in advance and feeding it into the faster memory tiers.

What Does This Mean for the HBM Market?

The skepticism from OpenAI and Intel executives suggests that CXL will not cannibalize HBM demand in the near term. Instead, the technology appears positioned to create a new memory tier that complements rather than replaces high-bandwidth memory. This scenario could leave the competitive landscape largely intact for Samsung Electronics and SK Hynix, which currently hold leading positions in the HBM market, even as CXL gains traction as an additional layer in the broader AI memory hierarchy.

The distinction is important: HBM remains essential for the compute-intensive phases of AI model execution, where data movement speed directly impacts performance. CXL, by contrast, excels at managing less frequently accessed data and expanding overall memory capacity without the cost premium of HBM.

Steps to Understanding CXL's Role in AI Infrastructure

• HBM vs. CXL Trade-offs: HBM provides extremely fast data transfer speeds essential for active AI model computation, while CXL offers larger capacity at lower cost but with significantly slower bandwidth, making it better suited for storing and managing less frequently accessed data.
• Memory Hierarchy Integration: CXL functions as an intermediate tier between fast DDR memory and slower solid-state storage, allowing data center operators to optimize cost and performance by routing different types of data to the appropriate memory tier based on access patterns.
• Inference vs. Training Workloads: CXL appears most valuable for AI inference tasks, where models process data less intensively than during training, making the bandwidth limitations less critical to overall system performance.
• Cost Implications: By reducing the amount of expensive HBM required in AI server deployments and using CXL for secondary data storage, data center operators can potentially lower overall infrastructure costs while maintaining acceptable performance levels.

The practical reality emerging from industry discussions is that CXL represents an evolution in how AI data centers manage memory resources, not a revolution that displaces existing technologies. Samsung and SK Hynix's investment in CXL production reflects their recognition that the AI infrastructure market will demand multiple memory solutions optimized for different purposes. As AI systems grow more complex and data centers scale to unprecedented sizes, the ability to intelligently distribute data across memory tiers with different speed and cost characteristics becomes increasingly valuable.

For Samsung and SK Hynix, this diversification strategy protects their market position even if CXL adoption accelerates. Rather than competing directly with HBM, CXL becomes another product category where these memory leaders can maintain their technological edge and market share. The 35.7% efficiency improvement SK Hynix achieved with its ITME architecture demonstrates that thoughtful integration of CXL into existing memory systems can deliver real performance gains, even if CXL alone cannot replace HBM's critical role in AI acceleration.