CXL 2.0: Shared RAM to Combat Shortages

A Surge in RAM Prices

In just one year, the cost of DDR5 memory has seen a spectacular increase, tripling compared to the previous year. This rise is largely due to the massive demand for HBM memory by GPUs dedicated to artificial intelligence, used in data centers. Memory manufacturers like Samsung, SK hynix, and Micron have had to redirect their production towards these more lucrative chips, leaving fewer resources for conventional DRAM. This situation has led to a cost pass-through to PC and smartphone manufacturers, who are forced to raise their prices.

The RAM shortage has been a hot topic since last winter, especially for AI-specialized data centers that consume enormous amounts of HBM memory to power the GPUs from Nvidia, AMD, and others. This demand has drained the production capacities of major manufacturers, who have had to focus on the most profitable chips. As a result, there is less conventional DRAM available, leading to a threefold increase in DDR5 prices since the beginning of 2025.

CXL: A Promising Technology for Data Centers

Compute Express Link (CXL) is an interconnect protocol that uses the same physical lanes as PCIe. Its principle is to pool memory among multiple servers, much like carpooling optimizes the use of parking spaces. Instead of dedicating a fixed amount of RAM to each server, which is often underutilized, CXL allows for the creation of a shared memory pool.

Three versions of CXL currently exist:

• CXL 1.0 (2019) allows for adding memory to a server via a PCIe slot.
• CXL 2.0 introduces switching and "memory pooling," enabling multiple servers to access the same memory reservoir, although each zone remains assigned to one host at a time.
• CXL 3.0 goes further by allowing true memory sharing between machines with cache coherence.

Current AMD Epyc and Intel Xeon processors already support CXL 2.0, and Amazon has announced that its Graviton5 is compatible with version 3.0.

A Powerful Consortium Behind CXL

CXL was initially developed by Intel in 2019 and is now managed by an open consortium that includes giants like Alibaba, Cisco, Dell, Google, HPE, Huawei, Meta, and Microsoft. AMD, NVIDIA, and Samsung joined the board in 2022. With over 250 members, the consortium released version 4.0 in November 2025, demonstrating massive industry support.

Companies like Astera Labs, Marvell, and Enfabrica are already working on "memory godboxes," dedicated memory enclosures that can hold up to 18 TB of DDR5 per chassis. Enfabrica promises to halve computing costs by connecting them to NVIDIA GB200 racks. NVIDIA acquired Enfabrica last September, recognizing the potential of this technology.

Samsung and Micron also offer CXL modules ranging from 128 to 256 GB. The most immediate use of these modules concerns the offloading of KV caches, which are the attention caches of language models. These caches can sometimes occupy more memory than the model itself. Google has already attempted to optimize these KV caches with a proprietary algorithm a few weeks ago.

Impact on the Consumer Market

Although CXL is primarily aimed at data centers, its effects will be felt in the consumer market. Currently, memory accounts for about 35% of the manufacturing cost of a PC, compared to 15 to 18% the previous year. Smartphone prices have also increased by 40 to 50% depending on the markets. Apple, for instance, accepted a 230% increase for the LPDDR5X in its iPhone 17 without negotiation, illustrating the shift in power dynamics.

If CXL allows data centers to pool their RAM, it could alleviate pressure on the consumer memory market. However, it will take several quarters before a significant impact is observed. In the meantime, DRAM prices are expected to rise by another 93 to 98% in the second quarter of 2026, with normalization hoped for as early as 2027. Thus, consumers should expect to pay more for their next DDR5 kit.