Once known mainly for controllers sold via OEM, Phison reports rapid growth in enterprise and AI storage, opens a European subsidiary in Eschborn, Germany, and promotes flash as a memory tier for local AI as shown on IT Press Tour.
Phison is trying to shed its image as the silent supplier behind other companies’ SSDs. At the 70th IT Press Tour, executives of the Taiwanese NAND flash specialist presented a company that now describes itself as an AI storage provider, with a new German subsidiary, a growing enterprise portfolio and a flash-based memory extension for AI PCs.
Michael Wu, president of Phison U.S., supplied the numbers. According to the company, revenue reached about 2.3 billion US dollars last year, and Phison expects 8 to 9 billion dollars this year. The first half alone brought in roughly 3.4 billion dollars, around 150 percent of the previous full year. Enterprise and AI-related business now accounts for close to 40 percent of revenue, against less than 5 percent a year earlier. “Phison has officially transformed into an AI storage company,” Wu said.
The starting point was a business that rarely stood in the spotlight. Founded about 25 years ago, Phison says roughly one in five client SSDs worldwide uses one of its controllers, often sold under another brand. Industrial and embedded products contribute a steady 16 percent, while retail has lost ground. About 80 percent of the budget goes into engineering, and the company employs more than 4,500 engineers.
Wu called the transformation painful. After USB sticks, SD cards and consumer SSDs came industrial and embedded designs, where defective products are not simply replaced and engineers debug on site. Enterprise is the latest step: Phison has opened an office in Colorado with about 30 employees and invests heavily in validation labs at its Taiwan headquarters. Its customers include server vendors, AI storage system vendors and chip ecosystem players. It also supplies boot drives, an area it says NAND makers treat as a low priority.
To demonstrate engineering depth, Phison also points to unusual projects. In its own blog, the company describes how it served as storage and controller partner for Lonestar Data Holdings’ Freedom Mission, which it calls the first hardware-based data center on the Moon. Pascari enterprise SSDs with customized controller firmware were validated under vacuum, vibration and extreme temperatures, since no repairs are possible on the lunar surface. The project won the DevOps Dozen award for Best DevOps Industry Implementation. These details come from Phison’s own account.
Wu sees AI limited at two ends: the energy going in and the data that must be stored afterwards. Phison expects token demand to grow roughly 50-fold within three to five years, while NAND supply is projected only to double. From this the company derives its argument that AI must move to the edge and that storage becomes the new compute, because cached results can be reused instead of recalculated.
Wu makes a related case in a March 2026 opinion piece for the Forbes Technology Council, a paid-membership platform, so the text reflects Phison’s own position. He argues that AI is held back less by compute and HBM bandwidth than by the speed at which data reaches the processors, which makes storage the new bottleneck for CIOs. To support the point, he cites Gartner, which predicts that organizations will abandon 60 percent of AI projects lacking AI-ready data through 2026, and an Accenture report in which 75 percent of executives name good-quality data as the most valuable ingredient for improving generative AI.
Wu also separates three phases of the AI life cycle with different storage demands. Training needs high throughput for parallel reads, smooth data movement across pipelines and tiers, and rapid scaling, with data volumes from petabytes to exabytes. Fine-tuning works with megabytes to terabytes, but depends on fast iteration, frequent checkpointing and versioning, and predictable performance. For inference, latency comes first: queries draw on embeddings, vector stores and token caches, and the same data is accessed repeatedly across distributed nodes. Wu adds that more inference is moving to the edge or near the edge, mainly for lower latency and better privacy.
Expansion to Europe
Europe is part of the plan. On 2 September, Phison announced Phison Electronics Europe GmbH in Eschborn, Germany, its first local entity. Ingvar Ersson, with more than 25 years in data storage, leads it as GM and President EMEA, reporting to CEO KS Pua. Antony Zukowski, previously at distributors such as TD SYNNEX and Arrow ECS France, is Regional Director EMEA. The company combines direct enterprise engagement with distributors, system integrators and partners, and ASBIS will bring the Pascari enterprise SSD portfolio to market through its channel. Ersson named media and entertainment, HPC, neoclouds, local government and education as first verticals. Phison’s CTO added that European customers welcome customized products, which larger suppliers often decline.
The channel is already taking shape. In January, Zstor GmbH of Kaarst secured the distribution rights for Phison’s Pascari series of enterprise NVMe SSDs in Europe. Zstor positions itself as a distributor and system integrator of storage solutions for enterprise customers across various industries. “As demand for high-performance storage continues to grow in Europe, Phison’s products—particularly the NVMe M.2 and E1.S drives—are well-suited for our enterprise customers,” explained Zstor CEO Helmut Kopel. A key focus of the partnership is aiDAPTIV+, a Phison technology that uses SSDs as GPU cache instead of traditional GPU VRAM and is designed to support the development and optimization of AI models.
The Pascari brand is about two and a half years old. At Computex, Phison showed four products, including the D206V, a QLC drive with up to 245 terabytes, and a liquid-cooling-ready E1 model. According to the company, the Pascari X200B ranked third in the IO500 benchmark. Large customers can also order tailored drives and on-site engineering support through the Imagine Plus program.
For client devices, Phison takes a different route. Its Adaptive technology adds flash as a cache tier next to GPU memory and DRAM, while middleware moves model weights and KV-cache data as needed. The drives run in pseudo-SLC mode and are designed for up to 100 drive writes per day, far above a regular TLC drive. In one example, a mixture-of-experts model needing about 24 gigabytes of memory ran on a 16-gigabyte system after up to half of the weights were moved to flash. The trade-off is speed, which Phison considers acceptable for overnight tasks. Citing Gartner, the company expects token usage to grow five- to thirtyfold this year as AI shifts from chat to agents, and 55 percent of PCs to be AI PCs. It announced a collaboration with Intel at Computex, and eight PC makers are evaluating the technology.
Dr. Jakob Jung is Editor-in-Chief of Security Storage and Channel Germany. He has been working in IT journalism for more than 20 years. His career includes Computer Reseller News, Heise Resale, Informationweek, Techtarget (storage and data center) and ChannelBiz. He also freelances for numerous IT publications, including Computerwoche, Channelpartner, IT-Business, Storage-Insider and ZDnet. His main topics are channel, storage, security, data center, ERP and CRM.
Contact via Mail: jakob.jung@security-storage-und-channel-germany.de