KIOXIA and Dell Put 157 PB of Flash in One 42U Rack and Shift the Data Center Bottleneck to Interconnect
KIOXIA America and Dell Technologies said on November 11 2026 that they have built a 157 PB object store in a single industry-standard 42U rack. The parts list is short. Sixteen Dell PowerEdge R7725xd servers, each holding 40 KIOXIA LC9 Series 245.76 TB SSDs, running Dell ObjectScale. That is 640 drives and about 157.3 PB of raw flash.
For anyone who plans storage refreshes, the number to hold onto is per drive, not per rack. One E3.L SSD now holds about 246 TB, which is many times what a single nearline hard drive stores. For read-heavy data, hard drives are no longer the only sane way to buy petabytes.
What the configuration actually is
Dell first qualified the LC9 drive for ObjectScale in August 2026, with 9.83 PB in one 2U server. At the time the rack figure was a projection: 20 servers, about 196.6 PB. The November build uses 16 servers, which fill 32U of the 42U rack and leave space for switching and power. That is a configuration someone could actually deploy, and it is the more useful number.
The LC9 is built on KIOXIA’s eighth-generation BiCS FLASH 3D QLC with a 32-die stack, on PCIe 5.0 and NVMe 2.0. It keeps working if a NAND die fails, supports Flexible Data Placement to cut write amplification, and is rated at 2.5 million hours MTTF. KIOXIA notes that MTTF is a statistical value, not a life guarantee. Security options include Sanitize Instant Erase, self-encrypting models and FIPS 140-3 validation, depending on country.
What the release leaves out is the part procurement teams need. There is no price, no availability date, no power draw, no benchmark. The power and cooling claims are qualitative: fewer systems, fewer racks, less cabling. QLC flash also has lower write endurance than TLC. The design fits read-heavy object data, such as training sets, data lakes and archives, much better than write-heavy transactional work.
The same week, compute and optics
Two other announcements from the same days fill in the rest of the rack.
The University of Alabama’s new statewide High Performance Computing Center, also built with Dell, runs PowerEdge R770 and XE7740 servers with NVIDIA RTX PRO 6000 Blackwell Server Edition and H200 NVL GPUs, NVIDIA Quantum-2 InfiniBand, and Dell PowerScale file storage. It is partly funded by a NIST award and by the State of Alabama. The release gives no node or GPU counts. It is a reference design for what a mid-sized research shop now buys: mixed GPU generations, InfiniBand, scale-out file storage.
At TSMC’s 2026 North America Technology Symposium in Santa Clara, Ayar Labs won the Demo of the Year award for an optical interconnect built into a chassis-level AI server reference design with Wiwynn. Two AI accelerators each carry eight TeraPHY optical engines and move more than 100 Tbps of bandwidth per accelerator. SuperNova external light sources in the standard ELSFP form factor drive the optics, and TSMC’s COUPE platform handles the photonic packaging. Ayar Labs has raised $650 million in equity in 2026.
What changes for IT buyers
Put the three together and the bottleneck moves. When one rack holds 157 PB and one accelerator wants 100 Tbps, capacity stops being the constraint. Moving data between storage, servers and accelerators becomes the hard part, and that is a networking budget line.
Three practical checks follow:
- Price flash per usable TB against your actual erasure-coding scheme, not raw capacity. Dell’s ObjectScale material describes 24+4 coding on these nodes.
- Model a failure. A 2U node holding nearly 10 PB raises the stakes of one bad chassis; ask for rebuild times on your workload, not the vendor’s.
- Check the uplinks before the drives. A dense flash rack behind a 100GbE fabric wastes most of what you paid for.
For the market side of the memory makers behind these drives, see technologies.org.