Samsung ships PM1763, a PCIe 6.0 SSD for AI training clusters

Samsung on July 8, 2026 began mass production of the PM1763, a PCIe 6.0 enterprise SSD for AI and HPC servers. The drive ships in 4TB, 8TB and 16TB capacities, with a quoted 28,400 MB/s read, 21,900 MB/s write, and a 1.4-second load time for a 40GB large language model. Per the Samsung announcement, PM1763 improves power efficiency 1.8x over PM1753 and pairs 9th-gen V-NAND with a 4-nanometer controller.

What PM1763 ships with on day one

The PM1763 is a single-sided, 2.5-inch enterprise SSD built around a PCIe 6.0 interface and Samsung's 9th-generation V-NAND. The 4-nanometer controller sits next to the NAND package, and the drive is optimized for direct-to-chip liquid cooling, which keeps the SSD at peak throughput under sustained AI training and inference workloads. Samsung ships three capacities on day one: 4-terabyte, 8TB, and 16TB. The 16TB configuration is the headline SKU and is the one Samsung positions against AI customers running frontier model training and inference clusters.

The performance numbers Samsung quotes are competitive with the best enterprise SSDs on the market. Sequential read is 28,400 megabytes per second, sequential write is 21,900 MB/s, and the drive is rated at more than 2 times the performance of the prior PM1753 generation. The single most useful benchmark for AI customers is the 40-gigabyte large language model transfer, which Samsung says the PM1763 completes in approximately 1.4 seconds. The same workload on the prior generation takes roughly three times longer, which is the difference between an SSD that disappears into the data path and one that introduces a visible training-stage bottleneck.

Power efficiency is the second leg of the announcement. PM1763 improves power efficiency by 1.8 times compared to its predecessor, which Samsung frames as a datacenter operating-cost win rather than a device-level spec. For a hyperscaler running tens of thousands of NVMe drives in a single training cluster, a 1.8x efficiency improvement on the storage layer is roughly equivalent to a free generation of free cooling capacity. The press release frames this in the same sentence as the AI training cost story, which is the right framing for buyers who are pricing out their next AI infrastructure build today.

The third feature is security. PM1763 supports post-quantum cryptography algorithms designed to protect against future quantum computing threats, and it supports the TEE Device Interface Security Protocol, a virtualized-environment standard for protecting data pathways between host and device. Both features target regulated enterprise buyers who need to show that their storage layer can survive a quantum threat horizon and that their virtualized NVMe fabric meets the attestation requirements that enterprise AI buyers now demand. Samsung does not break out pricing, and the press release is silent on which hyperscaler customers have placed volume orders.

The quote inside the announcement is from Jangseok Choi, Vice President and Head of Memory Product Planning at Samsung Electronics, who frames PM1763 as a key solution that enables customers to efficiently scale memory capacity and optimize AI operations. The framing is consistent with the broader memory-side pitch that AI customers now need to plan memory and storage together, and that the storage layer is now an active participant in the training cost model rather than a passive last-mile buffer.

Why a PCIe 6.0 SSD matters for AI infrastructure in 2026

PCIe 6.0 doubles the per-lane bandwidth of PCIe 5.0, and PM1763 is one of the first enterprise SSDs to put the standard into volume production. The jump matters for AI customers because the bottleneck on a modern training run is the gap between GPU HBM and the storage layer that holds the dataset, the checkpoint files, and the model shards. PCIe 5.0 SSDs were the first generation that could keep up with the per-GPU bandwidth budget on a frontier model training cluster. PCIe 6.0 SSDs are the first generation that can keep up with the next two generations of GPU silicon, which are the systems that will start deploying at scale in late 2026 and 2027.

The Micron and Anthropic HBM and storage supply line deal announced in June framed memory and storage as a joint planning surface for enterprise AI buyers. PM1763 is the storage-side complement to that argument. Where Micron and Anthropic agreed on HBM co-design and a multi-year memory supply line, Samsung is now committing to a PCIe 6.0 storage roadmap that gives the same buyers a stable endpoint for the storage layer. The combined message to enterprise AI buyers is that the supplier side of the AI memory and storage stack is now being planned as a single system rather than as two independent product lines. The practical result is that an enterprise AI buyer who engages Samsung, Micron, and Anthropic in 2026 will see a coherent capacity roadmap from HBM to NVMe, with the PCIe 6.0 generation arriving just in time to support the Blackwell Ultra and Rubin generation of GPU silicon that is now landing in production data centers.

The storage-side bottleneck is also where the AI training cost story gets interesting. A training run that spends 8 percent of its wall-clock time waiting on storage is paying for a memory tier that costs roughly the same as the GPU tier, and the waiting time is a pure overhead. PM1763's quoted 40-gigabyte LLM transfer in 1.4 seconds is the spec that moves that overhead from a noticeable training-stage tax into the noise floor, which is the threshold hyperscalers care about. The wider story is that the storage tier is becoming a first-class participant in the AI training cost model rather than a passive last-mile buffer, and Samsung's PM1763 is the most visible signal so far that the storage tier is now where the next generation of supplier-side AI infrastructure investment is landing.

The broader AI infrastructure landscape, including the GPU economics and the memory tier that decide whether an organization can afford an AI rollout, is documented in the AI Infrastructure in 2026 resource page. The full storage tier is the next layer of that stack that buyers will need to plan, and the PCIe 6.0 generation is the first generation where storage is positioned to keep pace with GPU silicon rather than trailing it by one generation.

Where PM1763 leaves the storage market

The immediate competitive question is which hyperscaler customers have placed volume orders for PM1763, and Samsung is not saying in the press release. The most likely first customers are the hyperscalers that Samsung already supplies with PM1753, which includes the major US and Asian cloud providers that build their own AI training clusters. Samsung's framing of the announcement as a mass-production event rather than a sample or a preview suggests the drive is shipping into revenue-generating systems this quarter, which is the same window that NVIDIA's Blackwell Ultra silicon is reaching volume production inside the same data centers.

The competitive set on the storage side includes Solidigm, Kioxia, Western Digital, and Micron, all of whom have PCIe 5.0 SSDs in volume production and PCIe 6.0 SSDs in development. The first mover advantage on PCIe 6.0 mass production is meaningful for hyperscaler buyers because the qualification cycle for a new SSD generation inside a hyperscaler data center is typically six to nine months, and Samsung is now several months ahead of the competition on volume shipments. The combined effect is that PM1763 is positioned to set the storage tier for the next generation of AI training clusters, with the second-mover competitors forced to either match Samsung's price or differentiate on a secondary feature such as power efficiency or security.

The security story is the second angle worth watching. Samsung's framing of post-quantum cryptography support on PM1763 is the first volume production commitment to PQC on an enterprise SSD from a major NAND vendor, and it sets a baseline that the rest of the industry will have to match. The TDISP support is the second half of the same story and is the standard that enterprise AI buyers will increasingly require for any virtualized NVMe fabric that handles sensitive model weights or training data. The combined PQC + TDISP story is the right answer for regulated enterprise buyers, and it gives Samsung a clean differentiation against the storage competitors that are still on the pre-quantum security stack.

The broader memory and storage supplier landscape, including the HBM, NAND, and DRAM planning that decide whether an organization can afford an AI rollout, is the same supplier-side stack that the AI Infrastructure in 2026 resource page documents. PM1763 is the storage-tier anchor for the next generation of that stack, and the timing is the right one for hyperscaler buyers who are finalizing their 2026 and 2027 AI infrastructure capacity plans in the third quarter of this year.