cd /news/ai-infrastructure/pcie-6-0-ssds-have-arrived-but-today… · home topics ai-infrastructure article
[ARTICLE · art-96671] src=mlq.ai ↗ pub= topic=ai-infrastructure verified=true sentiment=· neutral

PCIe 6.0 SSDs have arrived—but today’s buyers are AI data centers

PCIe 6.0 SSDs have arrived, but today's buyers are AI data centers, with Micron's 9650 entering mass production on February 12, 2026, and Samsung's PM1763 announced on July 8, both enterprise drives offering roughly 28 GB/s sequential read performance. Consumer platforms supporting PCIe 6.0 SSDs remain unavailable, with Phison's X3 controller not sampling until December 2026 and Silicon Motion's client-grade Gen6 storage planned for late 2027, according to Silicon Motion's June 2026 roadmap comments.

read6 min views1 publishedAug 14, 2026
PCIe 6.0 SSDs have arrived—but today’s buyers are AI data centers
Image: Mlq (auto-discovered)
  • PCIe 6.0 is a finalized standard, not merely a roadmap: it raises the raw link rate to 64 GT/s and uses PAM4 signaling, FLIT encoding, FEC and CRC to maintain reliability. [1] - Micron’s 9650 and Samsung’s PM1763 are the clearest commercially available Gen6 enterprise drives. Both use PCIe 6.0 x4 links and advertise roughly 28 GB/s sequential read performance. [2][3] - Phison’s X3 controller is scheduled to sample in December 2026 and ship in volume in mid-2027. Silicon Motion has discussed its SM8466 enterprise platform for 2026 but plans client-grade Gen6 storage for late 2027. [4][5] - The first buyers are hyperscalers, AI infrastructure operators and HPC deployments. Consumer platforms supporting PCIe 6.0 SSDs remain unavailable, according to Silicon Motion’s June 2026 roadmap comments. [5]

PCIe 6.0 SSDs are finally becoming real products, but the market that can use them today is narrow. Micron’s 9650 entered mass production on February 12, 2026, and Samsung announced mass production of the PM1763 on July 8. Both are enterprise drives designed for AI and high-performance computing systems, not retail desktop upgrades. [2][3]

The distinction matters because several other PCIe 6.0 announcements remain controller demonstrations, sampling plans or infrastructure components. Phison has shown a controller that is not scheduled to sample until December 2026. Silicon Motion has discussed an enterprise platform for 2026 but has pushed client-grade Gen6 storage to late 2027. Marvell’s public PCIe 6.0 announcement covers a switch and retimers, not an announced SSD controller. [4][5][6]

The standard is ready; the ecosystem took longer #

PCI-SIG released the PCI Express 6.0 specification in 2022. The standard doubles the PCIe 5.0 data rate to 64.0 GT/s and supports up to 256 GB/s of bidirectional bandwidth over a 16-lane link. A four-lane SSD connection has a theoretical one-direction bandwidth of about 32 GB/s before protocol overhead, which helps explain why current Gen6 drives top out near 28 GB/s rather than the headline 64 GT/s figure. [1]

PCIe 6.0 also changes the signaling and error-control machinery. It uses four-level pulse-amplitude modulation, or PAM4, to carry more data per signal cycle, along with fixed-size FLIT encoding, lightweight forward error correction and cyclic-redundancy checks. Those additions help compensate for the higher error rate associated with PAM4, but they increase the engineering burden for host systems, retimers, switches, connectors and SSD controllers. [1]

The interface standard is therefore settled, while the surrounding implementation remains uneven. PCI-SIG’s specification page identifies Revision 6.4 as the latest draft incorporating the PCIe 6.0 base specification, errata and approved engineering-change notices. Storage products also depend on separate NVMe and Open Compute Project requirements. Samsung lists the PM1763 as compliant with NVMe 2.1 and OCP 2.6; NVMe 2.3 was ratified in August 2025. [1][7]

Two enterprise drives are ahead of the controller pipeline #

Micron’s 9650 is the first clearly documented Gen6 data-center SSD to reach mass production. Micron lists sequential read performance of up to 28,000 MB/s, sequential writes of up to 14,000 MB/s, random reads of 5.5 million IOPS and random writes of 900,000 IOPS. The drive comes in E1.S and E3.S versions, with the E1.S design available in a liquid-cooling configuration. Micron’s product brief specifies a maximum power state of 25 watts. [2][8]

Samsung’s PM1763 is also in mass production. Its global product page lists a PCIe 6.0 x4 interface, sequential read and write speeds of up to 28,400 MB/s and 21,000 MB/s, random read and write performance of up to 6.8 million and 950,000 IOPS, and capacities listed as high as 61.44 TB. It lists 2.5-inch, E1.S and E3.S form factors and describes the drive as optimized for liquid-cooled server architectures. Samsung’s July 8 Korean launch release instead highlighted 4 TB, 8 TB and 16 TB configurations, with the 16 TB version rated for up to 21,900 MB/s sequential writes. [3][9]

That discrepancy appears to reflect different regional or product-line listings, so the exact capacity and write-speed availability should be confirmed with Samsung or a system vendor before publication as a purchasing claim. The broader point is clearer: Gen6 drives are entering enterprise qualification and production, but they are not yet standardized retail products. Vendor maximums also are not directly comparable without matching NAND, queue depth, endurance class, firmware and test conditions. [2][3]

Controllers still define the next wave #

Phison’s X3 is a 16-channel PCIe 6.0 x4 controller intended for enterprise storage. Phison demonstrated reference designs in E3.S and E1.S form factors, with claimed sequential read and write throughput of up to 28 GB/s, random performance of up to 6.8 million IOPS and capacity support reaching 2 petabytes per drive. The company says customer sampling is planned for December 2026, followed by volume shipments in mid-2027. Its stated power-efficiency target is 4 GB/s per watt, or about 7 watts at the claimed peak throughput. [4]

Silicon Motion’s SM8466 is another enterprise-focused platform. In a June 2026 interview, senior vice president Nelson Duann said the 16-channel controller was expected during 2026, while the company’s client PCIe 6.0 platform was planned for the end of 2027. Silicon Motion’s formal GTC materials describe the SM8466 as a MonTitan enterprise controller for near-GPU and AI-compute storage, but do not by themselves establish a shipping PCIe 6.0 drive. [5][10]

Marvell belongs in the PCIe 6.0 infrastructure discussion, but its March 2026 announcement was for the Structera S 60260, a 260-lane PCIe 6.0 switch, alongside Alaska P retimers. Marvell said the switch would begin customer sampling in the third quarter of 2026. That product can connect CPUs, accelerators, SSDs and CXL devices, but it is not itself an SSD controller or a drive. [6]

Why the extra bandwidth matters to AI storage #

AI systems move large model data sets, retrieval indexes, checkpoints and key-value caches between accelerators, memory and storage. A faster SSD link can reduce the time needed to stage data or refill a storage-backed cache, particularly when several drives operate in parallel. Micron positions the 9650 for AI training and inference, while Samsung describes the PM1763 as a drive for AI and HPC server environments. Those workload targets are vendor positioning, not proof that every AI application will run proportionally faster. [2][3]

PCIe 6.0 is most useful when the rest of the system can consume its bandwidth. That means a compatible CPU or accelerator platform, a Gen6-capable switch or retimer where required, suitable server wiring and software, and NAND capable of sustaining the controller. Without those pieces, a Gen6 SSD can operate below its advertised ceiling or fall back to an earlier PCIe generation. [1][5][6]

Power and cooling are part of the product decision. Micron specifies a 25-watt maximum power state and offers a liquid-cooled E1.S design. Samsung markets the PM1763 for liquid-cooled architectures, while Phison’s roughly 7-watt figure is a controller-platform efficiency claim rather than a universal power rating for a completed SSD. The first deployments will therefore favor centrally managed AI and HPC servers where the performance gain can justify qualification work and thermal infrastructure. [3][4][8]

Companies mentioned #

Further sources #

[[1] PCI-SIG describes PCIe 6.0’s 64 GT/s rate, up to 256 GB/s bidirectional x16 ban… ↗](https://pcisig.com/pci-express-6.0-specification)

[[2] Micron’s 9650 product page and product brief document February 2026 mass produc… ↗](https://www.micron.com/products/storage/ssd/data-center-ssd/9650-ssd)

[[3] Samsung’s PM1763 product page and July 8, 2026 announcement document mass produ… ↗](https://semiconductor.samsung.com/ssd/enterprise-ssd/pm1763/)

[4] Tom’s Hardware reported Phison’s X3 specifications, 2-petabyte capacity claim, … ↗

[5] Tom’s Hardware’s June 17, 2026 interview with Silicon Motion executive Nelson D… ↗

[6] Marvell’s March 17, 2026 announcement covers the 260-lane Structera S PCIe 6.0 … ↗+4 more The stories that matter, in one email. Free — unsubscribe anytime.

── more in #ai-infrastructure 4 stories · sorted by recency
── more on @micron 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/pcie-6-0-ssds-have-a…] indexed:0 read:6min 2026-08-14 ·