Innodisk Pushes Customizable SSDs and CXL Memory Into AI Servers

Innodisk announced on November 12, 2026 an expanded AI data-center portfolio that pairs enterprise SSDs rated at up to 14GB/s with DDR5 and CXL memory, targeting cloud service providers and neocloud operators running AI training, inference and high-performance computing.

The company's 3TS9-P, 4TS2-P and 5TS9-P SSD series are available for sampling and system-specific customization. Innodisk plans to showcase the full portfolio at SC26 in Chicago on November 16–19, 2026.

CC Wu, general manager of Innodisk's Flash Division, said the company is extending its industrial embedded-computing background into the AI data-center supply chain. According to Innodisk, customized firmware, security features, system-compatibility work and long-term supply commitments are intended to accelerate deployment, reduce integration risk and maintain infrastructure stability.

The promise of faster deployment carries a caveat: these drives are at the sampling stage, not volume shipping, and every claimed benefit depends on customers qualifying the products first.

What does the storage lineup actually offer?

Innodisk splits the portfolio into boot and operating-system drives on one side, and local storage for AI model weights, databases and inference datasets on the other.

The boot/OS range comprises the E1.S 5TS-P, 4TS2-P and 3TS9-P. All three offer customizable firmware, power-loss protection and end-to-end data-path protection.

For local data, the 5TS-P series is rated at up to 14GB/s sequential reads and 10GB/s sequential writes — roughly 14,000 megabytes per second, fast enough to move a full high-definition film in a fraction of a second. The E1.S model combines a PCIe Gen5 x4 interface with NVMe 2.0 and 3D eTLC flash, and spans capacities from 1.92TB to 7.68TB. Endurance is rated at one drive write per day, backed by a five-year warranty.

The E1.S product page also lists event logs, self-monitoring, online firmware updates and out-of-band management for large-scale deployments. The drive carries OCP Accepted recognition for the OCP Datacenter NVMe SSD Specification v2.0.

Thermal management is handled by a 15mm asymmetrical heatsink and thermal throttling, which Innodisk says sustains operation by limiting heat buildup and adjusting performance to prevent overheating.

How does the memory side address capacity limits?

Innodisk positions its memory range for long-context and agentic AI inference — workloads that increase the need for memory capacity and bandwidth while requiring lower latency. The lineup combines DDR5 server memory with CXL expansion modules.

The flagship DDR5 12800 Multiplexed Rank DIMM runs at 12,800 MT/s in capacities from 32GB to 128GB. Multiplexed registering clock driver and data buffer components enable simultaneous access to two ranks. Innodisk reports a 60% bandwidth increase over DDR5 8000 MT/s RDIMM.

The DDR5 8000 RDIMM is listed at 8,000 MT/s in capacities from 16GB to 64GB, using registered ECC memory. TVS diode and eFuse protection guard against power surges and overvoltage.

Not everything here is new. Innodisk's newsroom dates the MRDIMM introduction to July 30, 2026, and the DDR5 8000 RDIMM launch to September 29, 2026. The November announcement folds those earlier memory releases into the broader data-center portfolio rather than introducing fresh silicon.

The CXL Memory Module comes in 64GB, 96GB and 128GB versions, using an E3.S 2T form factor and a PCIe Gen5 x8 interface. Innodisk rates it at up to 32GB/s of additional bandwidth. Supporting CXL 1.1 and 2.0 through CXL.mem and CXL.io, it expands memory without additional DIMM slots and enables pooling across servers, allowing compute nodes to share memory resources.

What determines whether this matters commercially?

The portfolio gives Innodisk a broader AI-server offer spanning storage throughput, memory capacity and drive management. Its near-term business significance rests on customer qualification of the sampling SSDs — firmware customization, compatibility work and long-term supply support are what connect the company's industrial-memory background to its data-center ambitions.

Innodisk has not disclosed pricing, volume timelines, or named customers, so the market traction of the portfolio remains untested until qualification cycles conclude.