Kioxia and Dell cram 10 PB into slim 2RU server

Original Article ↗ Hacker News Discussion ↗

Ultradense flash storage is reaching new extremes, with Dell and Kioxia pairing 40 high-capacity NVMe SSDs in a 2U server to pack roughly 10 petabytes into a single chassis—at an estimated hardware cost well into the multimillion-dollar range. Commenters see this as technology aimed squarely at hyperscalers and niche high-end use cases, noting that endurance, data retention, thermals and PCIe bandwidth become major constraints at these capacities. Many hope such advances will eventually drive down SSD prices enough to displace hard drives for bulk storage, but note that current NAND supply issues and stagnant or rising consumer SSD prices make that future feel distant.

Pricing and Economics

  • Estimates put these 245–256 TB enterprise SSDs in the $15k–$50k range each, with recent anecdotes suggesting ~$400–$500/TB and sharp price increases over the last year.
  • Filling a 10 PB chassis is estimated at $600k–$2M in drive cost alone; some Dell configurator pricing suggests eight‑figure list prices for maxed-out systems, with the usual remark that “nobody pays sticker,” but discounts likely still leave it in the multi‑million range.
  • There’s frustration that such density is effectively limited to hyperscalers, defense, and high-end research, with jokes about needing to sell a house to afford one.

Consumer vs Enterprise and Trickle-Down

  • Multiple commenters dream of replacing large HDD NAS setups with single massive SSDs “in 20 years” or grabbing these cheap on secondary markets.
  • Others note enterprise SSDs are increasingly leased, not sold outright, and many are shredded for data-security reasons; only a minority of recycled drives are fit for resale.
  • Several people complain that SSD $/TB has stagnated or reversed for consumers; past builds with 4 TB NVMe were cheaper per TB than current 1 TB drives.

Reliability, Endurance, and Thermals

  • SSDs are described as “consumable” compared to HDDs; refurb experience suggests certain brands fail disproportionately, often after heavy cache duty in arrays.
  • Concerns are raised about data retention on very dense flash for archival purposes.
  • Each drive draws up to ~25 W; a full 40-drive server is ~1 kW, but even multiple such servers are seen as comparable to or lower than modern GPU racks in power density.

Form Factors and Connectivity

  • The drives’ E3.L (EDSFF) form factor interests homelab and portable-storage enthusiasts, who speculate about PCIe and USB-C adapter chains.
  • Some wish for high-capacity 3.5" SATA SSDs; others argue there’s no economic or technical reason now that M.2, U.2, and EDSFF exist and PCIe lanes are common.
  • PCIe 5.0 bandwidth is a bottleneck: with this many SSDs, networking is limited to a handful of 400 Gbps NICs; commenters look ahead to PCIe 7.0/8.0.

Use Cases, Scale, and Futures

  • Suggested use cases include high-density colo where space is the main constraint and massive backtesting datasets, though some argue HFT itself wouldn’t touch storage in hot paths.
  • One remark claims 10 PB could store lifetime records for an entire medium-sized country; another counters that such records are small and modest per-person telemetry remains limited.
  • There’s cautious excitement about “HDD killer” nearline SSDs and talk of a future where everyone runs large all-SSD NAS or “AI clouds” at home, tempered by skepticism over NAND costs and current supply constraints (e.g., Kioxia reportedly sold out of some products).

Space and Environmental Tangents

  • Long subthread debates orbital data centers/CDNs: critics cite radiation, limited chip rad-hardening, high power/heat, short satellite lifetimes, and difficulty of recycling hardware in space.
  • Others note that modern, smaller-node space-qualified SoCs do exist, with mitigation via redundancy and error correction, but still question whether any real-world problem is better solved in orbit than with terrestrial datacenters.
  • Some worry generally about the recyclability and repairability of ultra-dense electronics, though standard enterprise interconnects may at least help future reuse.