E1.S vs E3.S vs U.2: EDSFF NVMe Form Factors
If you are speccing storage for a current AI server, the short answer is EDSFF: E3.S for high-capacity and mixed-use drives, E1.S where you need the most drives in a dense compute node, and E1.L for hyper-dense QLC capacity. The legacy 2.5-inch U.2/U.3 form factor still ships in volume and populates most existing fleets, but new GB200-class designs are built around EDSFF because it cools hot PCIe Gen5 SSDs better and packs more terabytes per rack unit. Enterprise NVMe is the checkpoint and dataset tier of an AI datacenter — fast TLC for training and checkpoint I/O, dense QLC for read-heavy datasets — so the form factor you choose sets both thermal headroom and how many drives a chassis can hold.
This guide compares U.2/U.3, E1.S, E3.S and E1.L on density, thermals and use, and points to the right slot for GB200-class servers. Prices are indicative and as of 2026-08; for current SKUs and stock, see our enterprise NVMe SSD range.
Why EDSFF replaced the 2.5-inch drive
The 2.5-inch U.2 drive was designed for a spinning-disk world and inherits its airflow limits. As Gen5 SSDs draw more power and run hotter, that shape throttles under sustained AI I/O. EDSFF was purpose-built for flash: the modules present a better airflow profile, standardise on the PCIe/NVMe electrical interface, and let a chassis trade thickness for count or capacity. The result is more usable terabytes per rack unit and thermal margin for the newest controllers — which is why server vendors moved AI platforms onto it.
The four form factors at a glance
U.2 / U.3 — the 2.5-inch incumbent, typically 7mm or 15mm thick. U.3 adds a tri-mode connector that accepts NVMe, SAS or SATA on one backplane. Still widely stocked (Samsung PM9A3, Kioxia CM7, high-capacity Solidigm QLC), it is the safe choice for existing bays but the legacy path for new builds.
E1.S — a thin ruler (roughly 9.5–15mm) optimised for drive count and airflow in dense 1U compute and all-flash nodes. Drives like the Micron 7600 and Kioxia XD7P target this slot.
E3.S — a short module (commonly 7.5mm) that has become the mainstream AI-server slot, balancing capacity, endurance and thermals. The Micron 9550 and Samsung PM1743 ship here at up to ~14 GB/s.
E1.L — the long ruler built for maximum capacity per slot, home to hyper-dense QLC such as the Micron 6550 ION for read-intensive datasets.
Which form factor for GB200-class servers
New GB200-class AI servers are EDSFF-first. E3.S is the workhorse: it holds high-capacity Gen5 TLC for checkpointing and mixed-use I/O with the thermal headroom those drives need, and Micron’s 9550 in E3.S sits on the NVIDIA GB200 NVL72 vendor list. E1.S fits where a node prioritises raw drive count and tight airflow. For the dataset and inference tier, E1.L (or 15mm U.2) carries hyper-dense QLC — 30TB, 61TB and beyond — where cost-per-GB beats endurance. We help match the slot to the chassis across our storage catalog before you commit a BOM.
Capacity, endurance and price signals
Two commercial realities shape 2026 quotes. First, mainstream capacities (≤15.36TB TLC) carry public retail pricing, while 30TB+ TLC and 61–245TB QLC are frequently out-of-stock and sold on RFQ. Second, NAND contract prices rose roughly 95% over six months into mid-2026, so high-capacity SKUs age within weeks — quote per batch, not from a fixed list. TLC (1–3 DWPD) suits write-heavy training and checkpoints; QLC (~0.5–0.6 DWPD) suits read-intensive datasets. For sustained I/O and dataset storage we quote current enterprise NVMe SSD stock at RFQ.
| Form factor | Typical thickness | Density | Best for | Example drives |
|---|---|---|---|---|
| U.2 / U.3 | 7 / 15mm (2.5”) | Moderate | Existing fleets, tri-mode backplanes | Samsung PM9A3, Kioxia CM7 |
| E1.S | 9.5 / 15mm ruler | High drive count | Dense 1U compute, flash nodes | Micron 7600, Kioxia XD7P |
| E3.S | 7.5mm module | High capacity + thermals | GB200-class AI servers | Micron 9550, Samsung PM1743 |
| E1.L | Long ruler | Highest per-slot TB | Read-intensive QLC datasets | Micron 6550 ION, Solidigm D5-P5336 |
FAQ
What is EDSFF? EDSFF (Enterprise and Datacenter SSD Form Factor) is a family of ruler and short-module NVMe formats — E1.S, E1.L and E3.S/E3.L — designed for datacenter servers. Compared with the older 2.5-inch U.2/U.3 drive, EDSFF gives better airflow, higher rack density, and thermal headroom for hot Gen5 and Gen6 SSDs.
Which SSD form factor do GB200-class AI servers use? New GB200-class and other current AI servers are built around EDSFF — most commonly E3.S for capacity and mixed-use tiers, with E1.S in dense compute nodes. Micron’s 9550 in E3.S, for example, sits on the NVIDIA GB200 NVL72 vendor list. U.2/U.3 remains common in existing fleets but is the legacy slot for new designs.
What is the difference between U.2 and U.3? Both use the 2.5-inch drive shape. U.3 unifies SAS, SATA and NVMe on a single tri-mode backplane connector, so one bay can take different drive types. U.2 is NVMe-only wiring. U.3 backplanes accept U.2 drives, but not always the reverse — check host compatibility before mixing.
E1.S or E3.S — which should I choose? E1.S is a thin ruler for maximum drive count in dense 1U compute and flash nodes. E3.S is a slightly larger short module with more thermal and capacity headroom, and it has become the mainstream slot for high-capacity Gen5 drives in AI servers. Choose E1.S for count and airflow, E3.S for capacity and future-proofing.
What is E1.L used for? E1.L is the long ruler, built for maximum capacity per slot. It suits hyper-dense QLC drives — 30TB and up — for read-intensive AI datasets and inference storage, where per-slot terabytes and cost-per-GB matter more than peak endurance.
Bottom line
Build new AI servers around EDSFF — E3.S for capacity, E1.S for count, E1.L for QLC density — and keep U.2/U.3 for existing bays. Browse our storage range and request a quote for batch pricing on current Gen5 stock.