Investment in artificial intelligence (AI) data centres is expanding beyond high-bandwidth memory (HBM) to enterprise SSDs (eSSDs). Prices of high-capacity eSSDs in the 30TB to 64TB range for AI servers have risen by more than 50 percent recently as demand surges. The bottleneck is also shifting from production to post-process inspection.

Limits in supply capacity sit behind the rise in eSSD prices. According to TrendForce, Samsung Electronics and SK Hynix are responding by raising utilisation rates of eSSD lines based on 200-layer and above NAND flash. Increasing NAND output and shipping finished eSSDs on time are different issues, however. The higher the capacity, the longer pre-shipment inspection takes.

eSSD inspection is broadly divided into two stages: a burn-in test that runs devices for a long time in high-temperature conditions to filter out early defects, and a final test that checks read and write performance at actual interface speeds. As storage capacity rises from 32TB to 64TB and 128TB, the area that must be verified expands accordingly. That ultimately increases inspection time per product.

The spread of higher-layer NAND is also increasing the inspection burden. In a structure that stacks cells vertically to 200 layers and above, heat management and long-term reliability become more important. That means more conditions must be checked in burn-in tests. When interface speed rises to Gen5, heat generated during operation also increases.

Ultimately, reducing inspection time requires improving parallel processing so multiple products can be tested at once. That depends on the equipment’s number of channels and processing bandwidth. That is why replacing post-process inspection equipment is directly linked to shipment volume.

Equipment generation issues also overlap. NeoSem product specifications show that inspection equipment supporting PCIe 4.0 and equipment supporting PCIe 5.0 (Gen5) differ in the bandwidth they can handle. If Gen5 or higher eSSDs are tested with existing equipment, the equipment side hits a speed limit first. That is the backdrop to moving to new equipment that can run multi-channel final tests and high-temperature burn-in at the same time.

Data centres are also turning over faster. A McKinsey report said that as AI servers become more densely packed, replacement cycles for data centre servers and storage have shortened to about 2 to 3 years from 4 to 5 years. That means more storage must be replaced within the same period. The report projected that more than 70 percent of global data centre investment by 2030 will be concentrated on AI workloads. That also increases inspection volumes that must be handled in post-processing.

CXL 2.0 shifts inspection targets from chips to connections

New areas of inspection are also opening up, separate from replacement demand. Compute Express Link (CXL) is an interconnect standard that allows multiple servers to share memory. From CXL 2.0, it adds memory pooling (sharing) and switching functions. As a result, inspection targets expand from the operation of individual memory devices to connection behaviour between multiple devices.

Existing DDR5 DRAM modules or SSD testers have difficulty verifying this area. The issue is not whether memory operates normally, but whether errors occur as multiple servers share memory. That is why dedicated CXL testers are needed.

According to the industry, semiconductor post-process inspection equipment company NeoSem made the world’s first initial shipment of CXL 2.0 mass-production inspection equipment to a domestic memory company. The equipment includes PCIe 5.0-based high-speed bandwidth, functions to check memory sharing and switching malfunctions, and burn-in testing.

Demand-side evidence is also emerging. Microsoft is reported to have begun deploying commercial CXL memory on Azure by combining Intel Xeon 6 processors with an Astera Labs controller. It is understood to have cut total cost of ownership (TCO) by 15 to 20 percent in AI inference and retrieval-augmented generation (RAG) workloads. CXL has moved beyond the verification stage and is beginning to be used as a tool for actual cost reduction. In measurement equipment, Anritsu released high-speed interface evaluation equipment in June supporting CXL 2.0 and CXL 3.x.

For South Korean materials, parts and equipment companies, this is a period in which two streams of demand are opening simultaneously. Replacement of ageing equipment is already under way, while dedicated CXL lines are a newly forming market. In the short term, replacement of ageing equipment driven by a shift to PCIe 5.0 is cited. In the medium term, competition to secure an early lead in next-generation testers that extend to CXL 3.0 is cited as a variable. An industry official said, "Orders for inspection equipment are tied to memory makers’ line expansion schedules and depend on hyperscalers’ investment execution schedules, so the trend of commercial CXL 2.0 deployment is important."