The performance race among AI servers has long centered on GPU and memory bandwidth. But now that HBM4 has been implemented in NVIDIA's Vera Rubin platform, pushing memory supply to GPUs to 2.8TB per second, a new problem is emerging: the storage side responsible for feeding training data and inference weight files to those GPUs is starting to become the bottleneck for the entire system.
The sequential read speed achievable by PCIe 5.0-generation enterprise SSDs topped out at around 14,500 MB/s with Samsung's own previous-generation PM1753. Meanwhile, the granularity and frequency of data demanded by GPUs keep expanding in proportion to model size. If storage transfer speeds fail to keep pace with GPU processing speeds, expensive accelerators end up sitting idle. The move to PCIe 6.0 is the means of closing this gap. On July 8, 2026, Samsung brought its first mass-produced product in this category, the "PM1763," to market.
Numbers Enabled by 9th-Generation V-NAND and a 4nm Controller
Two newly developed components form the foundation of the PM1763: NAND based on Samsung's 9th-generation V-NAND, and a controller—an in-house chip newly developed using a 4nm process. Both represent a complete overhaul from the previous generation, affecting both performance and power consumption.
The pinnacle of performance is demonstrated by the 16TB model: sequential read of 28,400 MB/s and write of 21,900 MB/s. Both figures come from the company's internal benchmarks, with conditions undisclosed. Samsung states this represents more than double the performance of the previous-generation PM1753. As a concrete application example, the company cites the ability to transfer a 40GB large language model in approximately 1.4 seconds. However, this transfer time is based on the company's internal measurements and has not been independently verified by a third party.
The capacity lineup consists of three models: 4TB, 8TB, and 16TB. They comply with NVMe 2.1 and OCP 2.6, meeting the procurement standards required by hyperscale data centers. Compliance with the OCP (Open Compute Project) standard effectively serves as a gateway condition for large-scale procurement, since major hyperscalers such as Meta, Microsoft, and Google have incorporated it into their adoption criteria.
Cooling Design and Security Features for AI Clusters
The thermal design of the PM1763 is built around liquid-cooled servers. It supports "Direct-to-Chip (D2C)" cooling, which applies coolant directly to the chip, allowing sustained peak performance even during prolonged high-load operation. Behind the more than 1.8x improvement in energy efficiency (as claimed by the company) compared to the previous generation lies not only reduced power consumption from the 4nm controller but also its combination with this cooling mechanism. In high-density AI clusters, air cooling can no longer keep up with heat dissipation demands, making liquid-cooling support less of an added value and increasingly a procurement requirement.
On the security front, support for post-quantum cryptography (PQC) algorithms has been added. This aims to protect stored data even if future quantum computers reach a level capable of breaking current encryption. The drive also supports TDISP (TEE Device Interface Security Protocol), which cryptographically protects data paths in virtualized environments. Now that AI model training data and weight files have come to represent the core of corporate trade secrets, strengthening security at the storage level directly ties into customers' procurement requirements. PQC support reflects a design that anticipates future regulatory trends, and it's a factor that will be valued in government and financial-sector procurement as well.
Positioning for Vera Rubin and the Competitive Landscape
Samsung explicitly states that the PM1763 is optimized for NVIDIA's Vera Rubin servers. Vera Rubin is the platform NVIDIA positions as its next-generation AI infrastructure, and the PM1763 is expected to be adopted as a storage component within that ecosystem. This appeal targeting a "specific platform" carries more weight with procurement decision-makers than selling as a general-purpose product. The structure is such that if a data center operator chooses NVIDIA's stack, Samsung's storage naturally becomes part of the equation.
Competitors already exist in this space. Micron has been mass-producing its PCIe 6.0 SSD, the "9650," since early 2026, achieving sequential read speeds of up to approximately 28 GB/s and write speeds of about 14 GB/s. While the two companies are roughly on par in sequential read speed, Samsung's write speed of 21,900 MB/s exceeds Micron's approximately 14 GB/s. That said, there are aspects where Micron appears to be ahead in independent benchmarks and real-world operational data, so which company holds the advantage in actual deployment remains undetermined.
Third-party PCIe 6.0 products using the Phison X3 controller have also begun appearing on the market, and the range of options is expected to expand through the latter half of 2026. As a third player joins the Samsung-Micron duopoly, the leverage in procurement negotiations is likely to shift toward data center operators. For hyperscalers prioritizing supply diversification, this is a welcome development.
Pricing and the Path to Adoption
While mass production has been announced, shipping prices for the PM1763 have not yet been disclosed. Enterprise SSDs—including the previous-generation PM1753—often have prices negotiated on a model-by-model basis between companies, and it is not unusual for no public price to exist. At this stage, without clear pricing, it's difficult to assess the real hurdles to procurement. Price competitiveness will likely be the final variable in differentiating the PM1763 from Micron's offering.
What to watch next are adoption reports from major cloud vendors and data center integrators. Given that Samsung has explicitly named Vera Rubin as a target platform, whether the drive materializes in NVIDIA DGX systems or related architectures will significantly shape how the PM1763 is received in the market. Key indicators to watch going forward include when independent institutions release real-world performance comparisons against the Micron 9650, and when products equipped with the Phison X3 controller enter mass production—both of which will determine the competitive landscape of the PCIe 6.0 generation.