Shares of memory chip giant SK hynix (KRX: 000660) traded lower in Korea on Monday, falling 4.5% as of writing and touching an intraday loss of about 5%, the steepest single-day drop in two weeks. Aside from rising U.S. Treasury yields and elevated oil prices that reinforced risk aversion, another part of Monday's decline came from concerns that its "grandchild" company Solidigm might list in the United States, adding to worries about its complex ownership structure.

Media reports citing people familiar with the matter said Solidigm, the SK hynix unit focused on enterprise solid-state drives (enterprise SSDs), is pushing ahead with preparations for a listing. The company aims to complete the listing as early as next year, targeting about $15 billion in fundraising and a potential valuation of up to $150 billion. The expansion of agent applications represented by Muse and Astra is extending incremental AI infrastructure demand from model computing to task execution, efficient context management and massive-scale data storage, making enterprise SSDs an important area to watch in storage investment.

If it proceeds, the Solidigm listing would provide an independent valuation for the NAND flash and enterprise SSD company while attracting market capital on the back of the artificial intelligence infrastructure investment boom. Yet the potential listing news was read by investors as negative. The Korea Corporate Governance Forum criticized the Solidigm listing as further increasing SK Group's multi-layered shareholding relationships and deepening its pyramid ownership structure (SK Inc. to SK Square to SK hynix to AI Company to Solidigm), amounting to "exporting the Korea discount overseas." The forum had already urged SK hynix in August to abandon the Solidigm listing plan.

Investors also worry the Solidigm listing will lead to equity dilution. Once Solidigm lists independently, existing SK hynix shareholders will see their stake in this NAND asset, which is in the midst of an AI-driven supercycle, diluted, with profits diverted to outside shareholders. Jung In Yun, chief executive of Fibonacci Asset Management Global, said: "Rather than panic, I would choose to stay alert. A U.S. listing of Solidigm could unlock its value and provide funding for expansion, but SK hynix shareholders will give up part of Solidigm's future earnings." He added: "The key is the valuation and the use of the proceeds. Selling a small stake at an attractive price could create value. But massive dilution without a convincing case for return on investment would be concerning."

What exactly is Solidigm?

Solidigm is a U.S.-headquartered enterprise data storage company under SK hynix whose core business is NAND flash-based solid-state drives (SSDs) and related storage technology, with a focus on data centers, cloud computing and edge AI. It operates as an independently run subsidiary, headquartered in Rancho Cordova, California, and its website says it has 13 business locations worldwide and more than 2,000 employees. Its business foundation comes from Intel's former NAND flash and SSD operations. SK hynix announced in 2020 an acquisition of the related business for total consideration initially agreed at about $9 billion, completed the first-stage closing in December 2021 and set up Solidigm to take over product development, manufacturing and sales of the former Intel SSD business; the second-stage closing for the remaining NAND technology and manufacturing business was completed on March 27, 2025. As a result, Solidigm inherited Intel's long-accumulated enterprise storage technology, engineering team and customer relationships.

Specifically, it delivers complete enterprise SSD products to customers, and the value of a complete SSD comes from the coordination of flash media, controllers, firmware and system design. NAND is responsible for retaining data after power loss; controllers and firmware arrange data reads and writes, error correction, wear management and performance scheduling; enterprise products also need to meet requirements such as continuous operation, data integrity, write endurance and stable response times. Solidigm's capabilities therefore cover storage hardware, firmware and supporting software, with products optimized around customers' actual workloads. Solidigm's main business must be clearly distinguished from SK hynix's current core business, HBM. HBM is high-bandwidth DRAM that mainly provides high-speed data access at runtime for accelerators such as GPUs; Solidigm's core products belong to the NAND flash storage system, used to store large volumes of data and, under suitable software architectures, to carry part of reusable inference cache. The SK hynix group covers DRAM, HBM, NAND and SSDs, and Solidigm represents an important enterprise flash storage platform within it; the group's entire NAND storage chip business cannot all be attributed to Solidigm.

The more advanced and capable AI agents become, the more storage chips must scale up

The core change brought by Muse and Astra is that a single user instruction can launch a multi-stage, continuously running workflow. Meta disclosed that Muse runs on dedicated secure virtual machines and can execute tasks across applications; Astra strengthened computer operation, programming and complex professional work capabilities. A research or development task may continuously trigger information retrieval, file reading, code execution, model inference and result validation, and produce intermediate results that need to be retained. Extrapolating from the engineering architecture, GPUs and dedicated AI accelerators handle model computing, while high-performance CPUs handle browsers, virtual machines, tool execution and scheduling; enterprise knowledge bases, the data and indexes needed for retrieval-augmented generation (RAG), working files and audit records expand memory and persistent storage demand. As agent penetration rises, it is expected to drive both "computing power" and "data processing capability."

The second increment in storage demand comes from cache management needs generated by longer contexts and more concurrent tasks. In mainstream Transformer inference architectures, prefill processes input, decoding gradually generates output, and the key-value cache (KV Cache) stores reusable intermediate computation states. High-frequency data needed for active generation is carried by HBM, system DRAM handles buffering, and cache suitable for reuse can be placed in tiers across SSDs and shared flash according to access frequency and latency requirements, then preloaded back into memory. Nvidia's CMX architecture has explicitly proposed adding a flash layer for inference context between GPU memory and traditional shared storage. Its economic significance lies in expanding the capacity of context that can be retained and reused, reducing repeated computation and data waiting, thereby supporting more concurrent tasks. Enterprise SSDs thus gain new application space in participating in the inference process.

Solidigm's high-density products form a concrete connection with these needs. Its D5-P5336 has a maximum capacity of 122.88TB, uses QLC technology, and mainly targets large-capacity, read-intensive workloads such as data lakes and object storage. For data center operators, higher single-drive capacity helps reduce the number of devices needed to achieve the same capacity and optimizes rack space, power and cooling expenses; workloads requiring higher write performance or stricter response latency are handled by matching other SSD products and software configurations. From an investment perspective, Solidigm's strong growth opportunity comes from expanding AI data scale, upgrades in enterprise storage configurations and customers' continued pursuit of unit-capacity cost and system efficiency. Improvements in model efficiency can also lower the cost of completing tasks and attract more work into large AI inference systems; when the expansion of users and task scale exceeds the resource savings per task, demand for computing, storage, networking and power can continue to grow in tandem.