The 2026 Open Data Center Conference and inaugural Computing Power Expo (ODX 2026) was held in Beijing from Sept. 2 to 4. During the event, Samsung Semiconductor showcased three storage solutions for next-generation AI infrastructure: the PM1763, BM1773 and zNAND-O.

Jay Hyun, executive vice president of Samsung Electronics and head of NAND Flash Planning and Enablement, delivered a speech titled “Breaking Through the Memory Wall: Rebuilding the Data Path for Agentic AI.” He outlined the new demands that the development of agentic AI is placing on computing, memory and storage infrastructure.

Agentic AI Raises Pressure on Storage Systems

Artificial intelligence is evolving from generative AI toward agentic AI, Jay Hyun said. Agents can continuously iterate through planning, execution and reflection, while coordinating with multiple agents to complete complex tasks. This process generates large volumes of tokens, placing greater demands on computing, memory and storage infrastructure.

As agentic AI systems scale, the amount of KV Cache data is also growing. The speed at which KV Cache can be loaded affects GPU utilization. Its short lifecycle and large capacity requirements mean storage systems need high read bandwidth, low latency, and sufficient capacity and endurance.

PM1763 Focuses on PCIe 6.0 Performance

For high-performance AI systems, Samsung showcased the PM1763, an enterprise solid-state drive built around PCIe 6.0.

The product uses Samsung’s ninth-generation TLC V-NAND, a 4-nanometer controller and a PCIe 6.0 x4 interface, with capacities ranging from 4TB to 64TB. According to data released by Samsung, maximum sequential read and write speeds are 28,400MB/s and 21,000MB/s, respectively, while random read and write performance reaches up to 6.8 million IOPS and 950,000 IOPS.

These performance figures come from Samsung’s internal evaluations. Actual results may vary depending on system configuration and operating conditions.

BM1773 Offers Up to 256TB of Capacity

To address the capacity demands of large-scale AI workloads, Samsung also showcased the BM1773, an enterprise QLC solid-state drive.

The product uses ninth-generation QLC V-NAND and a 5-nanometer controller, with capacities ranging from 16TB to 256TB. The PM1763 and BM1773 focus on performance and capacity, respectively, serving the storage needs of different AI infrastructure workloads.

zNAND-O Explores Lowering Memory Costs for Edge AI

To address the pressure that expanding edge AI models are placing on memory costs, Samsung introduced the zNAND-O 3D storage solution.

According to Samsung, the core chip layer of zNAND-O uses SLC and combines 10th-generation BV-NAND with through-silicon via (TSV) technology. It can be integrated with a neural processing unit (NPU) in a package and interconnected through the UCIe standard interface, supporting both vertical and horizontal scaling.

Under the architecture proposed by Samsung, the operating system, applications, KV Cache and activation data remain in DRAM. Model weights, which require greater capacity and are accessed primarily for reading, are stored in zNAND-O, reducing the system’s reliance on high-capacity DRAM.

Samsung said zNAND-O can achieve 10 times the bit density of DRAM, while its read bandwidth and read energy efficiency can reach seven times those of conventional NAND flash. The company’s simulations showed that, compared with a conventional unified memory architecture (UMA) DRAM system, the solution could reduce memory costs while maintaining similar performance and accommodate models nearly twice as large.

These results currently rely mainly on technical data and simulation tests provided by Samsung. Actual performance and cost benefits will depend on specific system configurations and subsequent deployment.

Three Products Address Different Storage Needs

The three product categories Samsung showcased target the performance, capacity and cost requirements of AI infrastructure. The PM1763 focuses on high-performance storage using PCIe 6.0, the BM1773 on high-capacity enterprise storage, and zNAND-O on using edge AI systems to offload some of the storage burden from DRAM.

As agentic AI applications generate more model weights, intermediate states and KV Cache data, the efficiency of data transfers between memory, storage and computing resources is becoming a key factor affecting the performance and cost of AI systems.