Solidigm, a subsidiary of SK hynix, is adding a contract manufacturing site in Taiwan to its production network for data center SSDs. A product change notification (PCN) dated September 25 gives December 24, 2026 as the date by which customers should be ready to accept products made at the new site. It also says the existing sites will keep operating. Product designs and specifications will not change; the aim is to widen the system for supplying the same SSDs from multiple sites.
As AI infrastructure expands, buyers are also moving to secure SSDs when they need them. However, adding sites that assemble and manufacture SSDs is different from increasing production of NAND flash, the memory component itself, and the two affect different stages of supply. The significance of the new Taiwan site becomes clearer when you separate these stages and look at the relationship with Taiwanese server makers.
Specifications stay the same; manufacturing sites increase
Solidigm's PCN "0000048112-00" covers all of its data center SSDs. It adds a new ODM (a site entrusted with design and manufacturing) in Taiwan, with the goal of improving production flexibility and business continuity.
This is not a plan to replace existing facilities with the new site. Products will ship from both the current sites and the new Taiwan site.
The notice states explicitly that the products themselves are unchanged. The design, bill of materials and firmware are maintained, and the form factor and mechanical and electrical characteristics stay the same. So do the product names, product codes and published specifications.
The only change on the label is the first two digits of the serial number, which identify the manufacturing site.
In its customer guidance, Solidigm separates product requalification from administrative changes. Because the specifications do not change, it says requalification is not required. It does, however, ask customers to reflect the new first two serial-number digits in their internal records and incoming-inspection criteria.
In other words, even though performance does not need to be re-evaluated, customers need to be ready to track correctly which site made each SSD they receive.
The December 24 date is also not a confirmed shipping start date. The notice period is 90 days, and from that date customers may receive products made at the new site. The planned date could change depending on business and operational circumstances.
It is therefore more accurate to read December 24 as a guide for customers to prepare to accept such products after the end of the year, rather than as the start date for shipments from the new site.
Solidigm also says products made at each qualified site are interchangeable, and asks customers not to build procurement plans on the assumption of supply from a specific site.
What customers buy is an SSD that meets the qualified specifications. Even with the Taiwan site added, there will be no mechanism for specifying and purchasing only Taiwan-made products.
Ties to Taiwanese server makers
Taiwanese companies also play a major role in assembling the servers and systems deployed in AI data centers. NVIDIA's May 31 DSX announcement lists Foxconn, Pegatron and Quanta Cloud Technology (QCT) as participants. Wistron and Wiwynn are also developing systems that support DSX, the design and operations platform for AI data centers.
SSDs are chosen to fit the performance and operating conditions of these systems. It is not enough simply to install high-capacity drives. How efficiently a high-capacity SSD can be used depends on where data is placed and how writes are handled.
A joint study by Solidigm and Wiwynn is one concrete example. It tests a configuration that uses the "D7-PS1010," which has fast TLC NAND, as a cache, combined with the "D5-P5336" as high-capacity storage. A software layer called CSAL, placed between the two, optimizes writes to the storage SSD.
The background is the management burden specific to high-capacity SSDs. Internally, an SSD maps the logical locations specified by the host to physical locations on the NAND.
Enlarging this management unit saves memory for holding the mapping table. On the other hand, when small writes occur, the drive may need to read existing data, modify it and write it back in bulk. As a result, more data is written to the NAND than the host requested. This is write amplification.
Combining a cache with write control is meant to make high-capacity QLC SSDs easier to handle as part of the overall system.
However, the published results come from a specific sample configuration and do not guarantee the same performance on every server.
Expanding SSD production in Taiwan may also matter for working more closely with such system makers.
That said, the PCN does not disclose the name or location of the new contract manufacturer. Participation in DSX or the joint study with Wiwynn is not grounds to conclude that any of these companies is the new manufacturing partner.
AI inference growth is lifting SSD demand
In its August 7 announcement of NAND fab investment, SK hynix said that, in addition to the expansion of AI services, demand to store the KV cache used during inference is pushing up demand for enterprise SSDs.
Beyond the traditional use of storing large volumes of training data, SSDs are increasingly used to store and reuse results computed during inference.
The KV cache stores the Key and Value vectors an AI model computes partway through inference, reducing the burden of repeating the same calculations. The longer the context handled, the more data must be stored and reused.
In its March GTC announcement, Solidigm likewise explained that long contexts, continuous processing and repeated inference are increasing the amount of data handled on the inference side.
This does not mean SSDs run at the same speed as GPUs' high-speed memory.
What it means is that SSDs' role is growing within system designs that decide at which tier of memory and storage to place needed data and how to retrieve it when required. Securing enough SSD capacity and performance is becoming one of the conditions for expanding AI infrastructure as a whole.
The move to lock in supply capacity ahead of time is also showing up in actual contracts. On August 5, CoreWeave announced a multi-year strategic agreement with Solidigm, securing priority access to enterprise SSD supply capacity.
CoreWeave explained that storage has become a constraint on its capacity planning, citing rising demand and tight supply across the industry as the background to the deal.
Being able to procure GPUs alone does not guarantee that an AI cloud can expand as planned. Storage also has to be secured when it is needed. The agreement is an example of the companies involved describing that situation themselves.
However, contract volumes and amounts have not been disclosed, and there is no explanation that the Taiwan site being added will make SSDs for CoreWeave.
Increasing NAND output and expanding SSD supply are separate stages
Adding a Taiwan ODM is an effort to widen SSD production capacity. The CoreWeave agreement secures a supply allocation for a customer. M17 is an investment to strengthen the manufacturing base for NAND itself.
All of these bear on supply capacity, but the processes they target and their purposes differ.
Organizing official announcements as of October 7, 2026 by what each effort increases or secures gives the following picture.
| Measure and announcing party | Target | Disclosed timing and conditions |
|---|---|---|
| Solidigm's addition of a Taiwan ODM, notice of September 25 | Sites that manufacture SSDs with the same specifications | December 24, 2026 is the guide for customer readiness to accept products. Existing sites ship in parallel, and the schedule may change |
| CoreWeave–Solidigm multi-year agreement, announced August 5 | SSD supply capacity for CoreWeave | CoreWeave secures priority access to supply capacity. Volume and amount undisclosed |
| SK hynix's Cheongju M17 plan, announced August 7 | Fab base that manufactures NAND | Construction to start in February 2027, with the first cleanroom planned to open in December 2028 |
The dates in the table mark different milestones: customer readiness to accept products, the announcement of a contract, and the opening of a fab cleanroom. They are not figures that allow a like-for-like comparison of how soon supply will begin.
Adding a Taiwan ODM increases the number of sites that can make the same SSD, but it does not mean that production of the NAND those SSDs need will rise at the same time.
The investment decision for M17 is 19.1 trillion won. SK hynix says it will build the fab itself according to plan, while expanding the cleanroom and installing manufacturing equipment in stages according to demand.
For that reason, the December 2028 cleanroom target cannot simply be taken as the start date for NAND mass production or SSD shipments.
For the Taiwan site, meanwhile, neither the additional production volume nor the investment amount has been disclosed. The PCN also does not say whether it is a newly built plant or uses existing equipment, so it is not possible at this point to calculate how much it will boost output or how it will affect SSD prices.
Simply having more manufacturing sites should also be distinguished from spreading production across multiple regions.
The focus after the end of the year will be whether products made at the new site are supplied stably alongside those from existing sites, and what changes that brings to actual supply volumes and lead times.
If the necessary NAND can be secured and customers can set up systems to manage products arriving from multiple sites, making the same SSD at several sites could help strengthen the supply base supporting continued AI infrastructure growth.