In recent years, discussions on AI computing power have focused almost exclusively on GPUs. However, a frequently overlooked fact is that as large models enter the phases of long context windows, complex reasoning, and multi-agent systems, actual GPU utilization increasingly depends on whether model weights, KV Cache, and MoE experts can reach the correct compute nodes at the right time. Storage is moving from being a "peripheral device" into the real-time main path of token generation, which is the background for the emergence of AI SSDs.
AI SSD: From "Fast Storage" to "Efficient AI Usage"
In simple terms, an AI SSD is a solid-state storage device deeply optimized for AI large model training, inference, and AI workloads. Unlike ordinary SSDs, which primarily target general data storage, gaming, and office applications, the core logic of AI SSDs shifts from "how to store data quickly" to "how to enable AI systems to use data more efficiently," addressing data supply bottlenecks in AI scenarios through hardware-software co-design.
So, what are the essential differences between AI SSDs and ordinary SSDs? Regarding load scheduling, ordinary SSDs prioritize optimizing sequential read/write performance for large files, with limited optimization for4K small-block random read/write; whereas AI SSDs reconstruct firmware scheduling logic to specifically optimize 4K small-block random read/write and mixed sequential-random concurrent scenarios, significantly reducing latency jitter in AI inference. Regarding performance focus, ordinary SSDs emphasize peak sequential read/write speeds, where performance often drops sharply after SLC cache exhaustion; AI SSDs strengthen steady-state sustained performance in the absence of cache, while maximizing 4K random read IOPS to adapt to high-frequency small-access patterns in AI.
A more critical difference lies at the architectural level. Ordinary SSDs act merely as passive data storage devices, requiring data to traverse a complex chain of "SSD → CPU → Memory → GPU"; while AI SSDs introduce a storage-computing collaboration architecture. By partitioning the drive into dedicated AI Cache areas and standard storage areas, they achieve intelligent tiering of hot and cold data. Some products even feature intelligent prefetching and scheduling capabilities, allowing the SSD to participate in runtime data scheduling for AI inference. Furthermore, regarding thermal management under high load, endurance, and capacity/form factor adaptation, AI SSDs undergo specific optimizations for AI scenarios, such as tuning exclusive power and thermal strategies to support hours of uninterrupted high-load operation, and providing ultra-high-capacity versions to host massive model files and vector databases.
It should be noted that eSSDs (Enterprise SSDs) and AI SSDs are not parallel categories. eSSDs are product categories defined by application scenarios and reliability levels, emphasizing high endurance, power-loss protection, long-term stable operation, and data center compatibility. AI SSDs, however, are a concept defined by workload characteristics, emphasizing firmware, scheduling, and architectural optimizations for scenarios like AI training, inference, KV Cache, and vector retrieval.
In short, eSSDs answer "who uses this drive and is it reliable enough," while AI SSDs answer "does this drive understand AI workloads." Therefore, AI SSDs can be implemented based on eSSD form factors, or appear in edge and endpoint devices; not all eSSDs have undergone deep optimization for AI workloads. Among the vendors discussed later, many products appearing in enterprise eSSD form factors are also important carriers and basic forms of AI SSDs. Meanwhile, NAND flash and controller manufacturers jointly support the development of AI SSDs from different segments of the industrial chain.
Overview of Key AI SSD Players: International and Domestic Vendors
Having understood the definition and value of AI SSDs, the next question naturally arises: Who are the players in this track? Currently, the global AI SSD industrial chain has formed a complete division of labor ranging from NAND flash dies, controller chips, module brands, to system solutions, roughly divided into two major camps: international vendors and domestic vendors.
01｜International Vendors
International storage giants, leveraging long-term accumulation in fields like NAND flash, were the earliest and most systematic in laying out AI SSDs.
Samsung, as the global storage leader, has a full-stack AI storage layout covering HBM, SSD, CXL, and other product lines. At the GTC conference held in San Jose in March 2026, Samsung showcased HBM4, HBM4E, SOCAMM2, and PCIe 6.0 SSDs. Among them, HBM4 for NVIDIA Vera Rubin has entered mass production, while the PM1753 SSD demonstrated its capability to support inference workloads within the BlueField-4 STX reference architecture. Previously, Samsung launched a 256TB PCIe 6.0 SSD and planned a 512TB version, along with developing CMM-D storage products compatible with CXL 3.1 and PCIe 6.0.
SK hynix is a core supplier of HBM and AI storage, maintaining deep cooperation with NVIDIA. In December 2025, SK hynix Vice President Kim Chun-sung disclosed a next-generation storage solution codenamed "AI-N P." Based on the PCIe Gen 6 interface, this solution targets up to 25 million IOPS, approximately 8 to 10 times that of current mainstream high-performance enterprise SSDs, with initial samples planned for release by the end of 2026. SK hynix has begun R&D on second-generation products, aiming to push performance further to 100 million IOPS by the end of 2027. Additionally, the company is conducting joint proof-of-concept tests with NVIDIA on AI-N P, focusing on optimizing energy efficiency and speed during large-scale data throughput.
Micron is a supplier of data center storage and HBM, covering enterprise SSDs and advanced memory. In Q1 2026, Micron began mass shipping HBM4 for Vera Rubin and achieved mass production of PCIe Gen6 SSDs. Its data center-grade SSD portfolio includes three major series: flagship 9650, high-density 6600 ION, and mainstream 7600. The 9650 is optimized for agentic AI workloads on the NVIDIA BlueField-4 STX architecture, supporting up to 28GB/s sequential read throughput and 5.5 million random read IOPS; the 6600 ION series offers 122TB and 245TB versions.
Kioxia is a major NAND flash manufacturer that launched ultra-high IOPS SSDs for AI storage. On March 17, 2026, Kioxia announced in Tokyo the successful development of a new ultra-high IOPS SSD—the GP series. This product features XL-FLASH storage-class memory and targets the NVIDIA Storage-Next architecture, enabling GPUs to directly access high-speed flash as an extension of HBM. For long-context inference scenarios such as NVIDIA's context memory storage, Kioxia also launched the CM9 series PCIe 5.0 E3.S SSD, offering 25.6TB TLC capacity. In October 2026, Kioxia further announced adding direct liquid-cooling models to the CM9 series for SSD context caching scenarios, showcasing them at the OCP Global Summit in San Jose.
02｜Domestic Vendors
The landscape among domestic vendors presents a different picture: representative enterprises exist across all segments of the industrial chain, from NAND flash manufacturers to controller chips, module brands, and system solutions.
YMTC is a domestic 3D NAND flash manufacturer, whose NAND Flash series products are widely used in enterprise servers, data centers, consumer electronics, and other fields. Driven by AI computing power demand, enterprise SSDs are becoming a core growth direction. Currently, YMTC has launched multiple enterprise PCIe 5.0 SSDs including PE522, PE511, and PE501, all built upon its proprietary Xtacking4.0 technology.
T-Head Semiconductor, under Alibaba, started with the Zhenyue 510 enterprise controller to close the loop from chip design to cloud deployment. Currently, the Zhenyue 510 controller has been shipped in large volumes and adopted by multiple vendors such as Union Memory, Dera, and Biwin Storage, covering various capacity, performance, and price ranges, while continuously optimizing software stack and AI framework compatibility.
UNIS FLASH MEMORY is a core AI storage enterprise under Tsinghua Unigroup. Leveraging complete self-developed flash capabilities, it has built a full-stack AI storage product matrix covering "from cloud to edge." The new generation enterprise PCIe 5.0 SSD E5300 series is designed for AI data centers and core business scenarios. It adopts a technical solution combining domestic controllers with SK hynix 3D eTLC flash dies, fully exploiting PCIe 5.0 protocol transmission capabilities to achieve technological breakthroughs in sequential bandwidth, random performance, latency control, and energy efficiency. Relying on full-link reliability design, national cryptographic data security capabilities, and full-stack ecosystem adaptation, the E5300 supports operations such as large-scale databases, AI training/inference, virtualization, and real-time big data analysis.
Maxio, as an independent third-party SSD controller vendor, is actively expanding from the consumer market to the enterprise market. In March this year, Maxio launched the enterprise PCIe 5.0 controller chip MAP1803, supporting PCIe Gen 5.0×4 NVMe 2.0 interfaces. It employs a 16CH×8CE architecture with an IO rate of 2400 MT/s, achieving sequential read speeds up to 14.5 GB/s, random read speeds of 3.5M IOPS, and maximum capacity support up to 512 TB, targeting high-bandwidth, large-capacity, and high-reliability scenarios in large-scale data centers. In September this year, Maxio collaborated with Yinpu to launch an AI SSD solution, addressing the memory capacity wall issue in edge-side inference. This solution reduces memory usage by nearly half, lowers TTFT by 90%, increases TPS by 300%, and enables mainstream laptops to run large models.
Union Memory is one of the domestic providers of full-scenario solid-state storage solutions, covering enterprise, consumer, and embedded fields. In September 2026, Union Memory jointly released the high-performance KV Cache G3.5 solution for the Agentic AI era with Intel. Centered on Intel Xeon 6 processors, this solution builds a hierarchical KV Cache architecture based on the Union Memory UH812a enterprise PCIe 5.0 SSD, covering three tiers: GPU HBM, CPU DDR, and remote all-flash JBOF. Through end-to-end optimization using KVShrink, Intel QAT, SPDK, and RDMA, it provides scalable caching infrastructure for long-context, multi-turn interaction, and high-concurrency Agentic AI inference.
DapuStor is one of the few domestic vendors possessing full-stack self-research capabilities for data center enterprise "controller chip + firmware algorithm + module" with mass shipment. In August this year, DapuStor publicly displayed a 512TB QLC SSD for the first time at FMS 2026. In October of the same year, the 512TB SSD appeared at the 2026 OCP Global Summit, attracting industry attention. Two drives of this product can provide 1PB of flash capacity. In laboratory tests, its 128K sequential read performance reached 10.3GB/s, 128K random read bandwidth reached 4.02GB/s, and 4K random read performance reached 1.756M IOPS, providing high-capacity, high-performance storage support for scenarios such as large-scale AI datasets and data lakes.
Quanxing Technology is an AI storage vendor with its own manufacturing factory. Its wholly-owned smart manufacturing base, Quantian Technology, is located in Pingshan, Shenzhen, achieving a closed loop of the "design—R&D—packaging/testing—manufacturing" industrial chain. Products cover consumer, enterprise, and AI-related scenarios. Its "Tianyi AI Super-Memory Fusion Solution" features a "one card, dual use" core, allowing the SSD to serve as a high-performance PCIe 5.0 SSD while simultaneously acting as an "external VRAM pool" for GPUs. This design reduces reliance on expensive HBM by having the SSD handle massive KV Cache and long-context data, releasing onboard VRAM for immediate computation, thereby supporting larger models and lower inference costs on devices with limited computing power.
Pain Points, Trends, and Opportunities for Domestic Vendors in AI SSDs
After understanding the vendor landscape of AI SSDs, a more practical question arises: What challenges do AI SSDs currently face? Where lie the opportunities for domestic vendors?
01｜Major Pain Points of AI SSDs
The primary pain point is the bottleneck in performance and latency. AI inference has extremely high requirements for real-time data access, especially in frequent read/write scenarios for KV Cache. Traditional SSD latency struggles to meet GPU computing power demands, leading to "compute waiting for data." Traditional SSDs are designed for file and block data storage; their peak IOPS achieved under high queue depths does not equate to stable latency under low queue depth and fine-grained access conditions.
Secondly, there is a contradiction between capacity and cost. The surge in large model parameters and context cache data requires storage capacities of hundreds of terabytes or even higher. However, high-performance TLC SSDs are costly, while QLC, despite offering large capacity and low cost, faces issues of short write endurance and significant performance fluctuations. Under traditional unified memory architectures, placing all model weights in DRAM causes AI storage costs at the edge to skyrocket.
Third is the complexity of system collaboration. SSDs need to collaborate deeply with GPUs and CPUs to implement data prefetching, cache offloading, and intelligent scheduling. However, current multi-tier storage architectures are not yet mature, facing challenges in compatibility and scheduling efficiency. Traditional LBA layouts do not understand the semantic relationships between model layers, MoE experts, KV blocks, and their access order, potentially scattering interrelated data across different physical locations. To mitigate this, some vendors have begun letting SSDs actively participate in inference scheduling. For example, Quanxing achieves three-tier data scheduling between GPU VRAM, memory, and SSD via the AI Link algorithm, allowing KV Cache and MoE parameters to reside hierarchically based on heat.
Finally, there are risks related to supply chain and capacity. NAND flash capacity is concentrated among a few international manufacturers, with significant price volatility. In 2026, major NAND Flash manufacturers have added almost no new capacity. With sustained strong AI demand, the market exhibits a supply shortage throughout the year. Overseas giants are shifting capacity toward high-value-added products like HBM and server DRAM, squeezing traditional NAND capacity. New factory construction cycles last 2–3 years, causing new capacity release to lag behind the incremental storage demand generated by AI computing power.
02｜Industry Trends
Despite numerous pain points, industry trends have pointed the way forward.
The first trend is SSDs entering the real-time main path of inference. Mooncake by Moonshot AI and CMX by NVIDIA both point to a trend—inference state is becoming a first-class resource in AI infrastructure, and storage is beginning to enter the real-time main path of token generation. NVIDIA has established a dedicated Pod-level Flash context layer in the Vera Rubin infrastructure, indicating that SSDs are becoming resources that inference systems need to schedule directly.
The second trend is storage-computing collaboration becoming the mainstream architecture. From Quanxing's "Tianyi AI SSD" GPU external VRAM pool design, to Union Memory's joint KV Cache G3.5 hierarchical architecture with Intel, to more vendors starting to let SSDs participate in KV Cache tiering and offloading, storage is transitioning from passive hosting to active participation in AI inference data scheduling.
The third trend is accelerated domestic substitution. YMTC provides core media support for domestic AI SSDs. Domestic vendors continue to make breakthroughs in full-stack self-research of controller chips, firmware algorithms, and modules, with some products' performance already benchmarking against top international competitors.
The fourth trend is leading deployment of high-capacity QLC. In the QLC SSD field, vendors represented by DapuStor have already launched products of 122TB, 245TB, and even 512TB, compensating for media shortcomings through algorithm optimization and seizing the initiative in the AI cold data storage market.
Industry observers believe that behind these trends, opportunities for domestic vendors mainly manifest on three levels: first, accelerated domestic substitution; second, breakthroughs in full-stack self-research capabilities; and third, advantages in scenario-based innovation. Domestic vendors are closer to local AI customers and can perform deep firmware optimizations for scenarios such as KV Cache offloading, RAG vector databases, and edge inference.
How to Choose AI SSDs: Selection Criteria from Parameters to Delivery
Industry experts note that as AI SSD products proliferate, users often fall into the trap of "only looking at controller and Flash parameters" during selection. In reality, in true scenarios of high-load, long-duration AI operation, selection should comprehensively consider multiple dimensions.
Scenario match is the primary consideration. Different scenarios—enterprise AI training and inference, edge-side local large model deployment, and edge computing—have significant differences in SSD performance focus, capacity specifications, and form factor requirements. Training scenarios prioritize steady-state write bandwidth and endurance; inference scenarios prioritize 4K random read IOPS and latency consistency; edge scenarios have special requirements for power consumption, size, and capacity partitioning.
Controller and Flash specifications determine the upper limit of performance and the boundary of cost. The controller chip determines the SSD's performance ceiling and AI scenario adaptation capability, while the Flash type (TLC/QLC) directly impacts cost, lifespan, and capacity limits.
However, what often widens the gap between products is firmware and algorithm capability. Given the same controller and Flash, firmware scheduling logic and algorithm optimization capabilities are key. Whether the SSD optimizes for AI mixed read/write scenarios, supports intelligent KV Cache scheduling, and guarantees steady-state performance directly determines if the SSD can truly meet AI workload requirements.
An often overlooked but crucial aspect is manufacturing and delivery capability. Assuming fixed Flash and controllers, brands with their own manufacturing factories have relatively guaranteed competitiveness in supply, R&D, production, and pricing. Against the backdrop of tight NAND supply and capacity tilting towards enterprise grades, brands without in-house manufacturing capabilities face greater uncertainty in lead times and cost control. Among domestic vendors, Biwin Storage, TWSC, Longsys, and Quanxing Technology have all laid out their own factories or packaging/testing bases. For users, this means that amidst rapid iteration of AI projects and increasing customization needs, firmware adjustments, capacity configurations, and batch deliveries can be coordinated within a single system rather than through layered outsourcing. Such manufacturing capabilities are becoming an increasingly important source of certainty in AI SSD selection.
Finally, ecosystem adaptation and service. Enterprise deployments also involve compatibility with domestic CPUs and operating systems, as well as long-term supply commitments and technical support capabilities.
Conclusion
AI SSDs are moving from a "conceptual category" to a "rigid demand category." When HBM cannot fit everything and data movement becomes a bottleneck for large model inference, the entry of SSDs into the real-time data path is irreversible. For users, choosing an AI SSD is not just selecting a hard drive, but choosing a storage infrastructure capable of supporting long-term stable operation of AI businesses. In an industry context where Flash and controllers are gradually converging, brands with their own manufacturing factories, full-stack self-research capabilities, and scenario-based solution delivery capabilities are becoming increasingly important considerations in AI storage selection.