Dell and Kioxia Just Crammed 10 Petabytes Into a 2U Server — The End of Hard Drives Is Closer Than You Think
Trending on Hacker News today with over 100 points and 70 comments: Dell and Kioxia have achieved something that would've been unthinkable five years ago — 9.8 petabytes of all-flash storage in a single 2U server. The Dell PowerEdge R7725xd, packed with 40 of Kioxia's LC9 245.76 TB QLC NVMe SSDs, just rewrote the rules of data center storage.
To put that in perspective: a single rack of these servers can hold 196 petabytes. That's enough to store roughly 39 billion photos, or the entire Netflix catalog in 4K — multiple times over. And it's all flash, all NVMe, all fast.
Table of Contents
The Numbers That Changed Everything
Let's start with the raw specs, because they're genuinely jaw-dropping:
| Spec | Dell PowerEdge R7725xd |
|---|---|
| Form Factor | 2U rackmount |
| Total Storage | 9.8 PB (9,830 TB) |
| SSDs | 40x Kioxia LC9 245.76 TB |
| SSD Form Factor | E3.L (EDSFF) |
| Interface | NVMe PCIe Gen5 |
| CPU | AMD EPYC 9005 series |
| Networking | Up to 5x 400 Gbps NICs |
| Per-Rack Capacity | ~196 PB (20 servers) |
The networking spec is just as important as the storage. With up to 2 Tbps of network bandwidth per server (5x 400 Gbps), these boxes can actually move data fast enough to justify the capacity. A full rack with 20 servers could theoretically push 40 Tbps of storage traffic.
Inside the Kioxia LC9: A 245 TB SSD
The Kioxia LC9 is a beast of a drive. Key details:
- Capacity: 245.76 TB per drive
- NAND Type: QLC (Quad-Level Cell) — 4 bits per cell
- Form Factor: E3.L (EDSFF — Enterprise and Data Center Standard Form Factor)
- Interface: NVMe over PCIe Gen5
- Target: Read-intensive, high-capacity workloads
QLC NAND is the key enabler here. By storing 4 bits per cell instead of 3 (TLC) or 2 (MLC), QLC achieves the highest storage density per die. The trade-off is lower endurance (fewer write cycles) and slightly higher latency, but for read-heavy workloads like data lakes, AI training datasets, and cold storage, QLC is the sweet spot.
EDSFF (Enterprise and Data Center Standard Form Factor) is the successor to the traditional 2.5" U.2 drive form factor. E3.L is the "long" variant — taller and thinner than U.2, allowing better airflow and higher density in server chassis. Think of it as the industry's answer to "we need to pack more flash into less space."
The Dell PowerEdge R7725xd
Dell's R7725xd is built on the AMD EPYC 9005 platform (codename "Turin"), which provides the PCIe Gen5 lanes needed to actually drive 40 NVMe SSDs at full speed. The "xd" suffix indicates it's optimized for storage density.
Arun Narayanan, SVP of Compute and Networking at Dell, said the server "delivers the storage density and power efficiency our customers need to scale AI infrastructure without sacrificing performance."
The 5x 400 Gbps NIC capability is notable. Most servers ship with 1-2 network cards. Five 400G ports means this server was designed to be a storage node in a high-bandwidth cluster — exactly what AI training infrastructure demands.
The 256 TB SSD Arms Race
Kioxia isn't alone in this space. The 256 TB-class SSD market is heating up fast:
| Vendor | Product | Capacity | Status |
|---|---|---|---|
| Kioxia | LC9 | 245.76 TB | Shipping |
| Micron | 6600 ION | 256 TB | Shipping |
| Sandisk | UltraQLC SN670 | 256 TB | Shipping 2026 |
| SK Hynix | AIN D | 256 TB | Announced |
| Solidigm | (SK Hynix subsidiary) | 256 TB+ | Roadmap |
| Samsung | Nearline SSD | 256 TB+ | In development |
Scality has already confirmed it's working on supporting Samsung's upcoming nearline-class SSD, which they describe as an "HDD killer" with a roadmap extending to 1 PB per drive. When that arrives, a single 2U server could hold 40 petabytes.
What This Means for AI Infrastructure
This isn't just a storage story — it's an AI infrastructure story. Here's why:
AI Training Data Lakes
Modern AI models train on datasets measured in hundreds of terabytes. GPT-4's training data alone is estimated at 45+ TB. A single Dell R7725xd could store the training data for 200+ frontier AI models simultaneously, with enough bandwidth to serve that data to a GPU cluster at full speed.
Data Ingestion at Scale
Neville Ichhaporia, SVP and GM of Kioxia's SSD business unit, specifically mentioned "massive ingestion streams" and "data lakes" as primary use cases. For companies processing petabytes of telemetry, video, or sensor data daily, having all-flash storage means ingestion pipelines no longer bottleneck on disk I/O.
Backup and Disaster Recovery
Traditional backup infrastructure relies on tape or hard drives. With 10 PB in 2U, enterprises can keep weeks or months of backups on fast flash, dramatically reducing recovery time objectives (RTO). When a restore is needed, reading from NVMe is orders of magnitude faster than spinning tape.
Is the Hard Drive Finally Dead?
The "hard drives are dead" narrative has been around for over a decade, but this announcement brings it closer to reality than ever. Let's compare:
| Metric | 256 TB QLC SSD | 24 TB Enterprise HDD |
|---|---|---|
| Capacity per drive | 256 TB | 24 TB |
| Sequential Read | ~7 GB/s | ~270 MB/s |
| Random IOPS | ~500K+ | ~200 |
| Latency | ~100 μs | ~5-10 ms |
| Power (idle) | ~5W | ~8W |
| Vibration sensitivity | None | High |
| Cost per TB | ~$50-80 | ~$12-15 |
HDDs still win on cost per terabyte by roughly 4-5x. That's not nothing. For truly cold storage (data you rarely access), tape and hard drives remain more economical.
But the gap is narrowing. QLC NAND prices have dropped 60% in the past two years. Samsung's roadmap includes a 1 PB drive. At current trends, the cost crossover for "warm" storage (data accessed occasionally but needing fast retrieval) could happen within 2-3 years.
For AI infrastructure specifically, the argument for HDDs is already dead. When you're feeding data to a $100M GPU cluster, the cost of storage is a rounding error compared to the cost of keeping GPUs idle waiting for data. All-flash is the only sane choice.
The Cost Question
Dell hasn't published pricing for the R7725xd with the full 40-drive LC9 configuration. But we can estimate:
- A 245 TB enterprise SSD likely costs in the range of $12,000-20,000 per drive
- 40 drives = $480,000-800,000 for storage alone
- Add the AMD EPYC server chassis, networking, and Dell markup
- Estimated total: $600,000 - $1,000,000 per server
That sounds expensive, but let's do the math:
- At $800K for 9.8 PB, that's roughly $82 per TB
- A comparable HDD-based solution (417x 24TB drives) would need 10+ server chassis, more rack space, more power, more cooling
- TCO (total cost of ownership) including power, cooling, and floor space often favors flash even at higher upfront cost
Conclusion
Dell and Kioxia's 10 PB server isn't just an incremental improvement — it's a paradigm shift in data center storage. The implications are clear:
- AI infrastructure just got denser: Training data lakes that used to fill racks now fit in 2U
- HDDs are being cornered: Flash is winning everywhere except the coldest, cheapest storage tiers
- Network bandwidth is the new bottleneck: With 10 PB per server, the industry needs even faster interconnects
- The 1 PB SSD is coming: Samsung's roadmap suggests we'll see single drives with petabyte capacity within a few years
If you're building AI infrastructure, planning a data center refresh, or just keeping an eye on storage technology — this is the inflection point. The era of spinning disks for anything other than deep archival is ending.
Source: Blocks and Files — Kioxia and Dell cram 10 PB into slim 2RU server
Related: AI API Pricing Comparison 2026 | GGUF Format Guide: Running LLMs Locally