Samsung Electronics is expanding its AI memory and storage roadmap with new server SSDs, next-generation V-NAND and an in-development NAND architecture for on-device AI. At the Global Memory Innovation Forum 2026 in Shenzhen, Samsung Memory China CVP and CTO Kevin Yoon outlined technologies designed to address growing memory demands from AI inference and agentic workloads, including shared storage for key-value (KV) caches and higher-density memory alternatives for AI devices.
Samsung Targets AI Memory Bottlenecks as Agentic Workloads Grow
The transition from generative AI to agentic AI is changing the demands placed on memory and storage infrastructure. AI agents can execute multistep tasks, maintain context and interact with other agents, increasing the volume of intermediate data that systems must retain and retrieve. As workloads become more complex, memory capacity, bandwidth, power consumption and cost are becoming important considerations alongside GPU and accelerator performance.
At the fifth Global Memory Innovation Forum (GMIF) 2026, co-hosted by the Shenzhen Memory Industry Association and Peking University’s School of Integrated Circuits, Samsung presented a portfolio spanning enterprise SSDs, advanced NAND and an emerging memory architecture for AI devices. The summit attracted more than 3,000 registered professional attendees, according to the event announcement.
Kevin Yoon, CVP and CTO of Samsung Memory China at Samsung Electronics, argued that AI infrastructure needs memory and storage architectures designed around the requirements of agentic workloads rather than conventional generative AI alone.
PM1763 and BM1773 Address Shared AI Storage
Samsung outlined two server SSDs intended to support shared storage for AI key-value caches.
KV caches retain intermediate information generated during transformer-model inference. Reusing cached keys and values can reduce the need to recompute previously processed context, potentially improving inference efficiency. As models handle longer contexts and multiple concurrent requests, however, cache capacity can become a constraint.
Keeping all cache data in GPU memory may be expensive when workloads require large amounts of retained context. A shared storage tier could provide additional capacity, although performance depends on access latency, bandwidth, cache-management strategies and the workload’s reuse patterns.
Samsung’s PM1763 is positioned as the performance-focused option. The PCIe Gen6 SSD combines V-NAND TLC with an enhanced cooling system, and Samsung says sampling is planned for the fourth quarter of 2026.
The BM1773 targets capacity. Built on Samsung’s V9 QLC NAND, it spans announced capacities from 16TB to 256TB. Samsung describes the 256TB version as the world’s smallest SSD at that capacity.
The two products reflect different priorities in AI storage design. Performance-oriented SSDs can serve workloads sensitive to data access speed, while higher-capacity drives can help accommodate larger datasets and cache tiers. Neither capacity nor interface generation alone determines suitability for AI inference; real-world performance will also depend on latency, endurance, sustained throughput and software integration.
V10 V-NAND Pushes Beyond 400 Layers
At the NAND manufacturing level, Samsung introduced V10 BV-NAND, which it describes as the industry’s first V-NAND technology with more than 400 layers.
Samsung reports that V10 delivers 58% higher bit density, 33% faster input/output and 25% lower power consumption compared with V9. These are company-provided comparisons, and the announcement does not specify all testing conditions or measurement configurations.
The design uses an advanced bonding approach in which memory cells and peripheral circuits are manufactured separately before being bonded vertically. The objective is to increase storage density while improving performance and power efficiency.
Higher bit density can help manufacturers deliver greater capacity within a given footprint. That is particularly relevant as AI infrastructure expands and operators seek to increase storage capacity without proportionally increasing rack space, power requirements or cooling demands.
The broader industry challenge is balancing density with manufacturing complexity, reliability, endurance and cost. Samsung’s reported improvements will need to be assessed against production yields, commercial availability and independent performance measurements as the technology advances.
zNAND-O Targets AI Processing on Devices
Samsung also outlined zNAND-O, a next-generation NAND architecture under development for on-device AI. The company is developing four-layer and eight-layer versions designed to be co-packaged with a microprocessor unit (MPU) and connected through the Universal Chiplet Interconnect Express (UCIe) interface.
Samsung says the architecture delivers 10 times the bit density of DRAM, along with seven times the read bandwidth and read energy efficiency of NAND flash. Those comparisons refer to different memory technologies and should not be interpreted as demonstrating that NAND replaces DRAM across all workloads.
The underlying goal is to reduce the cost of accommodating increasingly large AI models on devices. Under traditional unified memory architectures, storing more model weights in DRAM can increase hardware costs and constrain the size of models that can run locally.
Samsung also described a simulation using Moonshot AI’s Kimi K2-1T model. According to the company, an NPU-optimized architecture using zNAND-O reduced memory cost to one-quarter at equivalent performance and accommodated a model nearly twice as large.
The results are simulation-based claims rather than evidence of a commercially available device achieving those outcomes. Independent benchmarks and details about the simulated hardware, workloads and cost assumptions will be important for assessing the architecture’s practical value.
Memory Architecture Becomes a Strategic AI Infrastructure Layer
Samsung’s announcements reflect a broader shift in AI infrastructure design: system performance increasingly depends on how efficiently processors access, retain and reuse data.
Enterprise SSDs, high-density NAND and specialized on-device memory address different parts of that challenge. For data centre operators, additional storage capacity could support larger cache tiers and more efficient inference pipelines. For device makers, denser memory architectures could expand local AI capabilities while managing hardware costs and energy consumption.
Samsung received the Outstanding AI Memory & Storage Industry Leadership Award at GMIF 2026, recognizing its contributions across DRAM, NAND, zNAND-O and CXL memory, according to the event announcement.
The next test is commercialization. PM1763 sampling is planned for Q4 2026, while zNAND-O remains in development. For these technologies to influence AI deployments at scale, customers will need validated performance, reliable manufacturing, compatible platforms and clear economic benefits.
Market Landscape
AI infrastructure is creating demand for memory and storage systems that can balance capacity, bandwidth, latency and energy consumption. Agentic workloads add complexity because models may need to retain and reuse intermediate information across extended tasks and multiple interactions.
Samsung’s portfolio addresses three layers of this challenge: PCIe Gen6 and high-capacity SSDs for server storage, higher-density V-NAND for enterprise capacity, and zNAND-O for emerging on-device AI architectures.
The competitive opportunity extends beyond NAND manufacturing to the wider AI memory hierarchy, including DRAM, CXL memory expansion, accelerator memory and storage software. Samsung’s reported performance improvements and simulation results remain company claims that require independent verification. Commercial adoption will depend on real-world benchmarks, system compatibility, manufacturing readiness and total cost of ownership.
Top Insights
- Samsung’s PM1763 and BM1773 SSDs target AI key-value cache storage, with one emphasizing performance and the other high capacity.
- The BM1773 uses V9 QLC NAND and offers announced capacities from 16TB to 256TB for enterprise storage deployments.
- Samsung reports that V10 V-NAND delivers 58% higher bit density, 33% faster I/O and 25% lower power consumption than V9.
- The in-development zNAND-O architecture targets on-device AI with higher memory density and a co-packaged design using the UCIe interface.
- Commercial impact will depend on independent performance validation, manufacturing readiness, platform compatibility and the economics of AI inference workloads.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
