Gate One: The Mapping Table Runs Out First
With 4KB-granularity logical-to-physical mapping, the table doubles every time capacity doubles. Past 30TB, the conventional ratio of roughly 1MB of mapping space per 1GB of capacity implies tens of gigabytes of table — beyond any realistic on-drive DRAM configuration. Two routes remain: enlarge the indirection unit, trading a 16KB or larger mapping granularity for a smaller table; or use tiered mapping with only hot regions resident. The first makes any write smaller than the indirection unit a read-modify-write, raising write amplification. The second adds a NAND read on every mapping miss, worsening tail latency.
The Choice Follows Access Granularity, Not the Capacity Number
The deciding factor is the smallest access unit of the workload. The Q7 J760Q and J760QD are planned for object storage, backup and archive, where writes arrive as whole objects. The Q9 J960Q and J960QD are planned for AI checkpointing and very large capacity arrays, where checkpoint flushes are equally large and sequential and then stay read-only for long periods. Both have access granularities far above 4KB, so the side effects of an enlarged indirection unit are diluted and that route is the sensible one. Metadata and logging tiers are the inverse — their writes are inherently small, the route does not apply, and X7 or X9 low-latency models cover them instead. Model names were frozen on 2026-07-22; subsequent changes to capacity, NAND and BOM are managed through SKU / PN / Revision / ECN, and every figure quoted here is a product-definition/EVT stage plan value.
Gate Two: The Failure Domain Grows With the Drive
Larger drives mean larger failure domains. When a 122.88TB drive fails, the array must move eight times the data a 15.36TB drive would require. If rebuild bandwidth is capped by the network or by throughput available from surviving members, the rebuild window stretches out considerably and the exposure period at reduced redundancy stretches with it. A drive cannot solve this alone, but its contribution is well defined: deliver stable and predictable sequential read bandwidth, use multiple namespaces to carve a large drive into smaller management units, and expose degradation trends early through telemetry so reactive rebuilds can become proactive migration. Multiple namespaces and telemetry are part of the shared feature baseline across the line, not a privilege of high-capacity models.
Gate Three: How QLC Write Performance Must Be Quoted
High capacity generally relies on QLC. Four bits per cell means sixteen threshold states instead of eight: programming needs finer verify stepping and takes longer, adjacent states sit closer together, the retention window narrows and cycling endurance falls. The direct consequence is that write performance has to be quoted in two separate terms — a short burst rate and a sustained rate at media steady state are not the same number, and reporting only the former carries no engineering meaning. Object storage, backup, archive and checkpoint writes are infrequent and large, so the steady-state rate is acceptable. Put QLC under a write-intensive workload and it meets the endurance ceiling and the steady-state write limit at the same time.
Two Boundary Figures at Opposite Ends of One Spectrum
The 122.88TB maximum single-drive capacity and the 120 DWPD maximum write endurance published for the line are two portfolio-level envelope figures, approached by different models in different series. Neither is the rated specification of any one series, and they cannot coexist in a single drive — capacity comes from more bits per cell, endurance from fewer, and the directions are directly opposed. Dividing 30 models into 10 series and 99 capacity SKU combinations is fundamentally a segmentation of that spectrum, letting each segment stay internally consistent within its own workload range.