A stylized motherboard shows glowing data streams flowing between memory modules, a storage drive, and a graphics card.
ORICO has announced the FlashMem FM9800, a PCIe 5.0 SSD that reserves dedicated space for local AI inference data alongside ordinary PC storage. Introduced on November 13, 2026, the forthcoming drive is designed around an additional software-managed storage tier—not a claim that NAND flash can magically become GPU memory.

The distinction matters: ORICO has described the architecture, but its announcement does not demonstrate an inference-performance improvement or establish broad Windows application compatibility.

A stylized motherboard shows glowing data streams flowing between memory modules, a storage drive, and a graphics card. What the FM9800 actually offers​

ORICO specifies an M.2 2280 drive using 3D TLC NAND, with a graphene thermal pad for the controller and flash components. Its two advertised capacities have separate allocations for conventional storage and AI caching:

Advertised modelStorage OS areaAI Cache area
1TB768GB85GB
2TB1,638GB128GB

These are ORICO’s published allocations, not independently measured usable capacities.

There is an unresolved specification question here. The listed areas total 853GB and 1,766GB, respectively—not the advertised capacities. ORICO does not explain the difference, so attributing it to reserved space, formatting or over-provisioning would be speculation. Buyers need a clearer capacity breakdown before treating either model as an ordinary 1TB or 2TB storage upgrade.

The software is the important part​

According to ORICO’s Chinese announcement, AI storage middleware and its Herdsman inference engine coordinate data across GPU memory, system memory and the SSD’s AI Cache. Selected model weights, context and inference-related data can occupy that flash tier when they do not need to remain in faster memory continuously. The two storage areas also receive separate wear-management and read/write scheduling policies.

That makes the proposal more specific than merely putting model files on a fast SSD. However, ORICO says results depend on the host, engine version, model, quantization and context length. Its announcement supplies neither application benchmarks nor a compatibility matrix identifying supported Windows versions and inference frameworks. A compatible physical slot should not be mistaken for confirmed AI-software support.

What PC users should establish first​

For anyone evaluating the FM9800 for local document processing, development or research, the sensible next step is a compatibility-and-results checklist:

  • Which operating systems, inference engines and model versions are supported?
  • Is Herdsman required, and how is the cache exposed to the operating system?
  • Does the configuration improve model capacity, response time, or both?
  • What happens during sustained inference and simultaneous ordinary storage activity?

These questions follow directly from ORICO’s workload-dependent design; the announcement does not yet answer them. Final endurance, warranty and performance specifications remain deferred, while a retail date and price are not stated.

The editorial takeaway is straightforward: assess this as a hardware-and-software system, not simply an SSD with an AI badge. The useful comparison will be the same model, context and host running with and without ORICO’s caching arrangement. Until that evidence arrives, the FM9800 is an interesting local AI storage design—not a demonstrated substitute for buying sufficient RAM or VRAM.

 

References

  1. ORICO Introduces PCIe 5.0 FlashMem FM9800 SSD for Local AI Workloads - XiaomiToday XiaomiToday 2026-11-16T19:36:05+00:00
  2. ORICO推出FlashMem AI SSD FM9800,开启本地AI存储新布局 orico.com.cn
  3. ORICO - USB Storage Adapters, Chargers, Hubs, and More orico.cc