Memory and Storage for SI, Part 6: The Capacity Drives

The largest SSDs now hold about 245TB each. Micron’s 245TB 6600 ION is available today, Kioxia’s LC9 reaches 245.76TB, and Sandisk said its 256TB UltraQLC drive would come in the U.2 form factor in the first half of 2026. All of them store four bits per cell, in quad-level cell (QLC) flash, and all three makers pitch them at SI clusters.

Part 4 and Part 5 covered the fast end of an SI cluster’s storage, which means drives tuned for small reads and a tier that holds the KV cache. This part is about the other end, where most of the bytes sit, and asks what these drives are for and what they give up to hold that much.

What makes a drive this large

Capacity comes from the flash die and how many fit in one package. Kioxia builds the LC9 from 32-die stacks of 2Tb QLC dies, made with the logic bonded straight onto the memory array (CMOS directly bonded to array, or CBA). Micron’s 6600 ION is built on its ninth-generation G9 QLC with six planes per die, and Solidigm’s D5-P5336 on 192-layer QLC, every one of them at four bits per cellA cell storing n bits needs 2 to the power n voltage levels: two for SLC, four for MLC, eight for TLC and sixteen for QLC..

Capacity per drive has doubled in steps. Solidigm introduced the 122TB D5-P5336 on November 13, 2024. Kioxia added its 245.76TB LC9 on July 21, 2025, beside a 122.88TB model, and Micron made the 245TB 6600 ION available on May 5, 2026.

Everpure (Pure Storage until February 23, 2026) is one maker that skips the drive format altogether and builds its own flash modules for its arrays, and they now reach up to 300TB, according to Computer Weekly. Flash Foundations Part 1 covers the bays and form factors that hold the standard drives.

Where they sit in an SI cluster

The makers describe the jobs in their own terms. For Kioxia, they are storing datasets for training large language models, and building the embeddings and vector databases that retrieval augmented generation (RAG) reads during inference. Sandisk aims its 256TB drive at data ingest, preparation, and fast SI data lakes. Micron frames it as a limit on the building: SI pipelines, analytics platforms, and object storage are growing faster than power, cooling, and space budgets allow.

Read together, those jobs suggest a read-heavy pattern. Data arrives in large pieces and then gets read many times, by training runs and by retrieval.

Demand has moved fast. In Counterpoint Research’s count, SI inference drove enterprise SSDs to 48% of global NAND shipments in Q2 2026, nearly double the 26% share a year earlier. That count covers every enterprise SSD, the fast ones included, so it is a measure of the whole server market rather than of capacity drives.

Part 5’s KV cache tier wants a different drive. Kioxia’s pick there is its CM9, a triple-level cell (TLC) drive rated at 3 drive writes per day (DWPD).

Bigger drives write in bigger pieces

An SSD finds data through a map from the addresses the host uses to the places the flash holds them, and Flash Foundations Part 6 explains why it needs one. Each map entry covers one indirection unit (IU). With 4-byte entries and a 4KB unit, the map is about 0.1% of the drive’s capacity, and the drive keeps it in DRAM. For a 122.88TB drive that is about 120GB of DRAM. With one entry per 32KB it is an eighth of that, about 15GB.

So larger drives map larger units. In Solidigm’s description, QLC drives use an indirection unit of 8KB to 64KB, one that grows with capacity. Its D5-P5336 uses 16KB on most models and 32KB on the 122.88TB model. Micron’s 6600 ION uses 4KB on the 30.72TB model and 16KB on the three larger ones.

Small writes carry that cost. A write smaller than the unit, or one that does not line up with it, makes the drive read the whole unit, change part of it, and write it all back. Solidigm calls that read-modify-write the main cause of higher write amplification, and notes that host software is generally tuned around 4KB writes.

Micron’s ratings put a number on it. The 30.72TB model has a 4KB unit and is rated at 0.3 DWPD for 4KB random writes, and the 16KB models at 0.075. That is a quarter as much. Across those two models the ratio matches the unit size over the write size, and 16KB / 4KB is 4. Writes that fill the unit don’t pay it, and the same 16KB models are rated at 0.3 DWPD for 16KB random writes and 1.0 for 128KB sequential writes.

Sandisk takes another route on the write path. It says its UltraQLC platform writes straight to QLC with no SLC buffer, and each write is safe against power loss on its first pass.

Endurance, read in petabytes

A DWPD figure looks small on these drives because it multiplies a very large capacity. Solidigm rates the D5-P5336 at 0.42 to 0.60 DWPD over five years, rising with capacity, and at 5.9PB to 134.3PB written across the same models. For the 122.88TB model, 122.88TB × 0.60 × 365 × 5 comes to about 134.6PB, close to the rated 134.3PB.

Those ratings carry their conditions. Solidigm rates the 16KB models with 16KB random writes and the 122.88TB model with 32KB random writes, so each figure assumes writes that fill the indirection unit, and the Micron numbers above show what smaller writes do to that budget. The same brief rates power-off retention at 3 months at 40°C, which is the figure to check before a drive goes on a shelf.

On client drives the worry about QLC wear is a separate question, and Flash Myths tests it on my own bench.

Against the hard drive: what the maker math shows

Micron’s brief sets a 245.76TB 6600 ION at 30W against a 44TB hard drive at 10W, which works out to 8.2TB per watt against 4.4. For 1EB of data Micron counts 4.7 racks of its drives at 212.3PB per rack against 31.5 racks of hard drives at 31.7PB, with 48 U.2 drives in each 2U server.

Solidigm’s figures for its 122TB drive are up to 84% less storage power in network-attached storage than hard drives paired with TLC flash, and up to 4PB per rack unit.

These comparisons set the SSD’s rated maximum power against a typical rating for the hard drive, and they measure the budgets Micron says are being outgrown: power, cooling, and floor space. The rack count is the most useful of them for planning floor space, and none of these documents gives a price per terabyte for either side, so a buyer has to bring that number to the comparison.

DirectFlash: the map moves to the array

Everpure moves most of the drive firmware into the array. Its DirectFlash Modules need no embedded controller doing the usual SSD work, because most of the tasks SSD firmware does move into the Purity operating system on the array’s own controllers, which maps the flash in every module. The modules also carry no DRAM of their own, and the array’s DRAM is shared across all of them.

That moves the map problem from the third section out of the drive. A module’s capacity no longer has to come with its own map memory, because the array holds one view of all its flash.

DirectFlash has buyers among the largest operators. Everpure’s first hyperscaler deal was with Meta in late 2024, in Computer Weekly’s report, and on August 10, 2026, Everpure announced a design win with a second top-five hyperscaler. It expects that deal to add significant revenue starting in its fiscal year 2028.

What to watch

Sandisk’s August 2025 release put its 256TB U.2 drive in the first half of 2026, and none of the sources saved for this page confirms it went on sale, so that is the first thing to check. Everpure expects revenue from its second hyperscaler deal starting in fiscal 2028, and the size of that order will show whether arrays that manage raw flash spread past the largest operators.

The trend I’d watch is the indirection unit. In Solidigm’s account it grows with capacity, and host software that writes 4KB at a time pays more for each step up. My read is that capacity drives keep the read-mostly jobs, datasets and data lakes, and that anything writing in small pieces stays on TLC drives like the one in Part 5.

The series turns next to the five-minute rule, which asks how often data has to be read to earn a place in faster memory.