According to a report citing people familiar with the matter, Solidigm, a subsidiary of SK hynix Inc. (SKHY), one of the world's largest memory chip makers, is considering an independent initial public offering in the US stock market as early as next year. The expansion of agent applications represented by Muse and Astra is extending the incremental demand for AI infrastructure from model computing to task execution, efficient context management and massive-scale data access, making enterprise SSDs a key area to watch in storage investment.
Against this backdrop, Solidigm, a unit of SK hynix Inc. (SKHY), is considering an IPO in the United States as early as 2027 and is in talks with potential advisers about a listing. The people said the world's second-largest DRAM/NAND memory chip maker is in discussions with potential advisers about a US listing of its NAND flash business subsidiary. The people asked not to be identified because the information is not public. Some of the people said the listing could value Solidigm at as much as a stunning $100 billion. They said discussions are ongoing and details such as timing and valuation could change. A US media outlet was first to report the details of the bank contacts and the listing timetable, citing people it did not immediately identify. A spokesperson for Solidigm declined to comment.
Solidigm's position in the industry happens to match the trend of AI investment spreading from accelerators to complete data center systems. Its core business involves NAND flash and enterprise SSDs, with products used in cloud computing, servers and data centers. Earlier, on August 5, CoreWeave announced a multi-year strategic agreement with Solidigm, securing priority supply of enterprise SSD capacity, and explicitly stated that storage has become a key constraint in AI platform capacity planning. This also means large AI cloud service providers are incorporating storage supply into long-term infrastructure construction plans to ensure that computing, networking and storage scale up in tandem.
For Solidigm, such cooperation helps improve demand visibility and provides a real customer basis for the independent valuation of its enterprise storage business; the announcement did not disclose the contract value or specific procurement capacity. In the era of AI inference, the core value chain of enterprise SSDs lies in delivering stable data access capabilities to customers through full and complete coordination of NAND, controllers, firmware and system validation: when expensive AI accelerators need continuous access to data, the value of storage is reflected simultaneously in capacity supply, computing utilization and the operating cost of the entire system.
In 2021, SK hynix Inc. (SKHY) acquired Intel's flash chip business and renamed it "Solidigm," and the company emerged as a result. According to its website, the company produces massive-data NAND storage products for data centers, some of which are only the size of a deck of cards yet offer up to 122TB of capacity. In August this year, the company announced an agreement with CoreWeave to sell enterprise solid-state drive storage capacity to the large AI cloud computing company, a so-called "neocloud" player, to support CoreWeave's AI cloud platform as a leading force in "AI neoclouds." According to its website, other customers include VAST Data, Dell Technologies and Chinese internet giant Tencent Holdings. It is understood that this storage chip leader, headquartered in Rancho Cordova, California, has 13 business locations worldwide, including Mexico, Canada and China, and more than 2,000 employees.
Storage chip components for AI data center server clusters remain the clearest supply bottleneck in the AI computing industry chain. Market research firm TrendForce expects server DRAM contract prices to rise by a cumulative 270% in 2026 and enterprise SSD prices to rise by a cumulative 235%; in 2027, HBM contract prices may still rise by 70% to 140%. These figures reflect the combined effect of AI computing expansion and storage price increases. TrendForce's latest estimate shows that DRAM and NAND together will account for 47% of major cloud service providers' capital expenditure in 2026 and 68% in 2027, driven by both higher procurement volumes and higher prices.
On the stock side, as of the US market close on September 25, 2026, based on end-2025 closing prices, in respective local currencies and excluding dividends, US storage chip leader Micron Technology Inc. (MU.US) has surged about 279.2% year to date; SK hynix Inc.'s (SKHY) Korea-listed shares have risen about 186.0% cumulatively.
What exactly is SK hynix's Solidigm?
Solidigm is an enterprise data storage company under SK hynix Inc. (SKHY) headquartered in the United States. Its core main business is NAND flash-based solid-state drives 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 official website discloses 13 business locations worldwide and more than 2,000 employees.
Its business foundation comes from Intel's former NAND flash and SSD business. SK hynix Inc. (SKHY) announced in 2020 that it would acquire the related business for total consideration initially agreed at about $9 billion. The first-stage closing was completed in December 2021, and Solidigm was established 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 is lost; controllers and firmware handle data read and write scheduling, 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 business capabilities therefore cover storage hardware, firmware and supporting software, and it optimizes products around customers' actual workloads.
Solidigm's main business must be clearly distinguished from SK hynix Inc.'s (SKHY) current main business, HBM. HBM is high-bandwidth DRAM and 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 preserve large-capacity data and, under appropriate software architecture, to carry part of reusable inference cache. The SK hynix Inc. (SKHY) group covers DRAM, HBM, NAND and SSD businesses, 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 AI agent performance becomes and the more work it does, the more storage chips must expand on a large scale
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 in a dedicated secure virtual machine and can execute tasks across applications; Astra strengthens computer operation, programming and complex professional work capabilities. A research or development task may continuously trigger data retrieval, file reading, code execution, model inference and result validation, and generate intermediate results that need to be retained.
From an engineering architecture perspective, 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 required for retrieval-augmented generation (RAG), work files and audit records expand memory and persistent storage demand. As agent penetration rises, it is therefore expected to drive both "computing capability" and "data processing capability."
The second source of incremental 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 in SSDs and shared flash according to access frequency and latency requirements, then loaded back into memory in advance. 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 the inference operation process.
Solidigm's high-density products form a concrete connection with the above demand. Its D5-P5336 reaches 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 optimize rack space, power supply and cooling expenses; workloads requiring higher write performance or stricter response latency are handled by other matching SSD products and software configurations.
From an investment perspective, Solidigm's strong growth opportunity comes from the expansion of AI data scale, upgrades in enterprise storage configuration, and customers' continued pursuit of unit capacity cost and system efficiency. Improved 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.
Comments