Enterprise Storage · GBA Technology · Published 2026-09-07 · Updated 2026-09-29

High-Capacity Storage 2026: Roles for 245TB QLC SSDs and 44TB HDDs

Single enterprise SSDs have reached 245.76 TB, and enterprise HDDs have reached 44 TB with HAMR. As capacity grows, the division of work between TLC, QLC and HDD depends more on write patterns, data temperature and the cost of recovery after a failure than on a single cost-per-terabyte figure. This article reviews the main high-capacity QLC SSDs and enterprise HDDs of 2026, the write characteristics of QLC, and how AI data can be tiered across media.

Key Takeaways

SSD capacity: The 245.76 TB Micron 6600 ION began shipping in May 2026; Kioxia LC9 also offers 245.76 TB; SanDisk has published a capacity roadmap to 256 TB and 512 TB.
HDD capacity: Seagate Mozaic 4+ reaches 44 TB with HAMR and is shipping to hyperscale cloud customers; WD is shipping 40 TB UltraSMR drives, with a 44 TB HAMR drive expected in the first half of 2027.
Write patterns: QLC endurance varies sharply with write pattern. For Micron 6600 ION it is 1.0 DWPD for 128KB sequential writes, 0.3 DWPD for 16KB random writes and only 0.075 DWPD for 4KB random writes.
Density: Micron says the 245.76 TB E3.L version can hold more than 3.9 PB per 1U.
Recovery cost: The larger each drive, the more data must be rebuilt after a failure; at 3 GB/s sequential write, the theoretical minimum time to fill a 245.76 TB SSD is about 23 hours.

High-Capacity SSD and HDD Overview

High-Capacity QLC SSDs

Model	NAND	Interface	Capacity	Form factor	Performance and endurance	Status
Micron 6600 ION	Micron G9 QLC	PCIe 5.0 x4	30.72 to 245.76 TB	E3.S (up to 122.88 TB), E3.L, U.2 15mm	13.7 GB/s sequential read, 3.0 GB/s write; 1.78M random read IOPS	Shipping
Kioxia LC9	BiCS FLASH gen 8 QLC (2Tb die)	PCIe 5.0 (single-port x4 / dual-port x2)	30.72 to 245.76 TB	2.5-inch, E3.S (up to 122.88 TB), E3.L	Up to 0.3 DWPD	Announced
SanDisk UltraQLC SN670	BiCS8 QLC	PCIe 5.0	30.72 to 122.88 TB; 256 TB version planned	U.2	13.7 GB/s sequential read	Available 122.88 TB
Samsung BM1743	7th-gen V-NAND QLC	PCIe 4.0 on U.2; PCIe 5.0 on E3.S	Up to 61.44 TB	U.2, E3.S	7.2 GB/s sequential read, 2.0 GB/s write; 0.26 DWPD	Shipping
Solidigm D5-P5336	QLC	PCIe 4.0	Up to 122.88 TB	U.2, E1.L	7.0 GB/s sequential read, 3.3 GB/s write (61.44 TB version)	Shipping
SK hynix PS1101	—	PCIe 5.0	245 TB class	—	—	Disclosed

Enterprise HDDs

Platform	Recording technology	Top capacity	Platter configuration	Status
Seagate Mozaic 4+	HAMR	44 TB	10 platters, more than 4 TB each	Shipping to hyperscale cloud customers
WD ePMR (UltraSMR)	ePMR with UltraSMR	40 TB	—	Shipping
WD HAMR	HAMR	44 TB	11 platters, 4 TB each	In customer qualification expected 1H 2027

Read-Mostly Data Needs a Different Design

Data of similar size can be accessed in entirely different ways; media choices should follow read/write ratios and recovery requirements rather than cost per terabyte alone.

A dataset reused repeatedly for training and a database log that is constantly rewritten may occupy similar capacity. The first cares about whether data arrives on time; the second must also sustain writes over a long period. The division of work between TLC and QLC becomes meaningful from this difference.

The purpose of storage tiering is not to label each dataset "cold" or "hot," but to separate read, rewrite and recovery requirements. Concentrated checkpoint writes during AI training also need to be evaluated separately from dataset reads, so that they are not hidden inside read averages.

TLC and QLC: Bit Density, IU and Write Patterns

QLC trades 4 bits per cell for capacity, and the cost falls on writes: write speed, write amplification and endurance all depend on the size and pattern of writes.

TLC stores 3 bits per cell and QLC stores 4, holding about one-third more data in the same area. High-capacity QLC SSDs usually use a larger indirection unit (IU); both Micron 6600 ION and Kioxia LC9 use 16KB. When a write is smaller than the IU, the controller must read, merge and rewrite the whole unit, increasing write amplification.

Micron 6600 ION's endurance specification shows this clearly:

128KB sequential writes: 1.0 DWPD;
16KB random writes: 0.3 DWPD;
4KB random writes: 0.075 DWPD.

The same SSD can differ in endurance by more than ten times depending on write pattern. Read and write performance is similarly asymmetric: Micron 6600 ION delivers 13.7 GB/s sequential read and 3.0 GB/s sequential write. QLC is therefore best suited to read-heavy data written in large sequential blocks; frequent small random writes, such as database logs and checkpoints, are better suited to TLC.

245TB-Class QLC SSDs

Competition in high-capacity SSDs has moved from 61.44 TB to 122.88 TB and on to 245.76 TB, with public roadmaps beyond 256 TB.

Micron 6600 ION: Uses Micron G9 QLC with capacities from 30.72 TB to 245.76 TB. The 245.76 TB version began shipping on May 5, 2026, in E3.L and U.2 15mm form factors; the 30.72 TB to 122.88 TB versions are also available in E3.S. Maximum power is 30W.
Kioxia LC9: Uses 2Tb BiCS FLASH QLC dies stacked 32 high into an 8 TB package, with CBA (CMOS directly Bonded to Array) technology. The 245.76 TB version comes in 2.5-inch and E3.L form factors, with an E3.S option at 122.88 TB; it supports PCIe 5.0 as single-port x4 or dual-port x2.
SanDisk UltraQLC SN670: Uses BiCS8 NAND, currently up to 122.88 TB. SanDisk's published capacity roadmap is 256 TB in 2026 and 512 TB in 2027, with 1 PB as a longer-term goal.
SK hynix PS1101: A 245 TB-class PCIe 5.0 high-capacity SSD that SK hynix disclosed at the Dell Technologies Forum in Seoul in October 2025.
Samsung BM1743: Uses 7th-generation V-NAND QLC, up to 61.44 TB; the U.2 version uses PCIe 4.0 and the E3.S version PCIe 5.0, with 0.26 DWPD endurance.
Solidigm D5-P5336: A PCIe 4.0 QLC SSD of up to 122.88 TB in U.2 and E1.L form factors.

Raw and Usable Capacity

SSDs are labeled with user-usable capacity; the difference between raw NAND and usable capacity is reserved for over-provisioning and error correction.

The 245.76 TB Micron 6600 ION has 256 TB of raw NAND, and 122.88 TB-class products have 128 TB of raw capacity. Over-provisioning affects write amplification and endurance. Capacity planning should start from usable capacity and then subtract data-protection overhead such as replicas or erasure coding.

The Benefits of Capacity and Concentrated Failure Risk

Larger drives need fewer slots, less power and less rack space, but concentrate more data in a single failure unit.

The density benefit is substantial. Micron says the 122.88 TB E3.S version can hold more than 2.4 PB per 1U, and the 245.76 TB E3.L version more than 3.9 PB. Micron also compares 720 bays: populated with 245.76 TB E3.L SSDs they reach 176.9 PB of raw capacity, versus 31.7 PB with 44 TB HDDs. These figures are calculated on the vendor's rack assumptions, and actual ratios depend on each data center's configuration.

Concentrated failure risk is the other side of the same trend. At Micron 6600 ION's 3.0 GB/s sequential write rate, the theoretical minimum time to fill 245.76 TB is about 23 hours; during a rebuild, the remaining devices must serve normal requests while also taking part in recovery, so actual time is usually longer. When evaluating high-capacity designs, measure throughput and tail latency both in normal operation and during rebuilds, and account for replica or erasure-coding overhead, spares and maintenance windows in usable capacity.

Enterprise HDDs: 44 TB HAMR and 40 TB UltraSMR

HDDs still support cold data and large data lakes on cost per terabyte and total capacity, and their capacity roadmap is being driven by HAMR.

Seagate: The Mozaic 4+ platform uses heat-assisted magnetic recording (HAMR) to reach 44 TB with 10 platters of more than 4 TB each, shipping in volume to two hyperscale cloud customers since March 2026. The previous Exos 30 TB and 32 TB drives deliver sustained transfer rates of about 270 MB/s.
Western Digital: Its 40 TB ePMR drive with UltraSMR began shipping in the quarter ended July 3, 2026. The 44 TB HAMR drive uses 11 platters of 4 TB each, is in customer qualification, and is expected in the first half of 2027. WD's published HAMR roadmap reaches 60 TB in 2028 and 100 TB in 2029.

SMR (shingled magnetic recording) increases density by overlapping tracks, so rewrites must also rewrite adjacent tracks, making heavy random rewrites slower than on CMR (conventional magnetic recording). HAMR instead briefly heats the recording medium with a laser to raise density without overlapping tracks, giving more predictable behavior in mixed read and write workloads.

Data Tiering: Dividing Work Between TLC, QLC and HDD

Data type	Access pattern	Typical media
Checkpoints, database logs, transactional data	Frequent writes, many small random writes	TLC SSD
KV cache and context memory	Low-latency, high-IOPS reads	TLC SSD (PCIe Gen5 / Gen6)
Training datasets, model repositories, vector databases	Read-heavy, writes mostly in large sequential blocks	QLC SSD
Data lakes, image and log archives	Large capacity, moderate to low access frequency	QLC SSD or HDD
Backups and cold data	Rarely read, cost per terabyte matters most	HDD

Cold data is not data that is never read, and warm data does not have a fixed access frequency. Tiers should be defined by actual recovery-time requirements: if cold data suddenly becomes a source for model training, the time needed to move it may matter more than the price difference between media. Tiering policies need to evolve with workloads.

TCO and Evaluation Framework

Total cost of ownership includes, beyond acquisition cost, power, cooling, slots, rack space, maintenance and replacement, as well as the service impact during rebuilds.

Dimension	Metrics
Normal reads	Sustained read bandwidth and tail latency on representative datasets
Concentrated writes	Read latency during write bursts such as checkpoints, and endurance calculated from actual write sizes
Rebuilds	Rebuild time after a drive failure, throughput and tail latency during the rebuild
Capacity	Business-usable capacity after replica or erasure-coding overhead
Density and energy	Capacity per 1U, power per terabyte, power ceiling per rack
Retention	Comparison of options under the same retention period and protection policy

Testing normal reads, concentrated writes and rebuilds on representative datasets while observing tail latency reveals problems that averages hide. The final design can mix TLC, QLC and HDD rather than treating one medium as a site-wide choice.

FAQ

What workloads suit QLC SSDs?: Read-heavy data written mostly in large sequential blocks, such as training datasets, model repositories, vector databases and data lakes. Frequent small random writes sharply reduce QLC endurance; Micron 6600 ION, for example, is rated at only 0.075 DWPD for 4KB random writes.
What is the largest enterprise SSD today?: Micron 6600 ION and Kioxia LC9 both offer 245.76 TB versions, with Micron 6600 ION shipping since May 2026. SanDisk has published a capacity roadmap to 256 TB and 512 TB.
What is the largest enterprise HDD today?: Seagate Mozaic 4+ reaches 44 TB with HAMR and is shipping to hyperscale cloud customers. WD is shipping 40 TB UltraSMR drives, with a 44 TB HAMR drive expected in the first half of 2027.
Will high-capacity SSDs replace HDDs?: The division of work between them is shifting. QLC SSDs lead on density, read performance and capacity per 1U, while HDDs keep an advantage in cost per terabyte and cold data storage. The right mix depends on access frequency, recovery-time requirements and total cost of ownership.

Sources

Last updated: 2026-09-29