[Digital Today reporter Chi-gyu Hwang (황치규)] Memory chip supply continues to lag demand as AI spreads. As big tech companies speed up efforts to find alternatives, CXL (Compute Express Link) technology is in focus for enabling servers in data centres to share memory.

In a typical server, memory chips are installed in fixed slots. One server cannot use memory left unused on a nearby server. It must use only the memory installed inside. CXL makes this possible. It means a server can draw on memory chips located elsewhere in the same data centre.

CXL is based on an open standard, which is managed by an organisation called the CXL Consortium. Members include Samsung Electronics, Nvidia, Intel and AMD, as well as cloud service companies such as Google and Alibaba.

CXL was introduced 7 years ago, but its spread has been slow until recently. A constraint has been that using CXL slows data transfers when a processor pulls data from memory to perform computations. In AI workloads, processors must fetch new data, and even small delays can slow the entire system.

Even so, as price rises worsen due to shortages of memory chips, companies are seeking various responses and appear to be increasing spending on CXL.

According to a recent report by The Information, Google has started deploying CXL technology in its data centres. Citing 2 Google employees, The Information reported this and said Google's move could influence other companies to take the same step.

Google is reported to have begun installing CXL controllers to manage traffic between server CPUs and large external memory pools. Google is also reviewing ways to integrate these memory pools more deeply into its systems. That would allow processors to access external memory faster, The Information reported, citing Google employees.

Even if Google has started deploying it in practice, CXL is still far from being established. The Information reported that industry experts say CXL still does not meet the level required by large cloud providers.

AMD released server chips supporting CXL in 2022 and Intel did so in 2023, but both products have seen only limited adoption in practice, The Information reported. Some also view the fact that development is not controlled by a single company as an obstacle to CXL adoption.

For CXL to be widely used, components that make up a data centre such as servers, CPUs, memory and network equipment must support the same standard. This is not something a company can decide in a few weeks, and coordinating it at the level of a large consortium could take much longer, The Information reported.

Even so, meaningful moves around CXL continue in the tech industry. Nvidia plans to support CXL 3.1 in its Vera CPU to be introduced late this year. Nvidia's move could be the largest real-world test around CXL, The Information reported.

Moves by South Korean companies are also taking clearer shape. Samsung Electronics, for its part, built CXL (Compute Express Link) infrastructure certified in 2024 by global open source solutions company Red Hat.

SK Telecom signed a memorandum of understanding at MWC 2026 being held in Barcelona, Spain, for joint development of a "CXL (Compute eXpress Link)-based next-generation AI data centre architecture" with Panmnesia. The two companies will apply CXL-based technology to expand the scope of resource interconnection from within servers to the rack unit level and shift to a structure that can flexibly combine CPU, GPU and memory. SK Hynix also completed customer certification last year for its CXL 2.0-based DRAM solution, the CMM (CXL Memory Module)-DDR5 96GB product.