Topicsstorage

Storage: Disks, Flash, and the Data an AI Model Eats

A sourced reference collection on the bottom of the memory hierarchy — why hard drives needed a laser to keep growing, what separates a data-centre SSD from the one in a laptop, why the storage interface had to be redesigned for a medium with no moving head, and why the accelerator now talks to the disk directly.

Assertions
17
Sources consulted
21
Read in full
7/21
Cited as evidence
7

21 sources sit behind this page — including any that arrive with a concept this page shares with another collection. 7 were retrieved and read in full, and only those can back an assertion. 14 were surfaced and deliberately set aside. Every one of them is named in the register below, with the reason in view. How we source this.

Background

Our own synthesis, written to orient you — not evidence. Every factual statement here is asserted and sourced further down this page.

Storage is the component people assume is solved. It is not, and the reason is visible in one number: hard drive areal density grew about 100% a year between 1996 and 2003, and about 10% a year between 2010 and 2016. The easy density ran out. What replaced it was a decade-long effort to put a laser inside a hard drive.

That effort is heat-assisted magnetic recording, and it shipped commercially in early 2024 after more than ten years of delay, requiring redesigned media, redesigned heads and new lasers. Seagate's Mozaic platform states 4TB or more of areal density per disk and drive capacities up to 44TB from a ten-disk stack, built on superlattice platinum-alloy media, a plasmonic writer with an integrated nanophotonic laser, a spintronic reader and a 7nm controller. A laser and a photonic waveguide inside a hard drive is not an incremental product update; it is a different machine in the same case.

Flash hit its own wall earlier and answered the same way everything else in these collections has answered: by going vertical. Planar NAND stopped scaling around 16 nm, and Samsung commercialised the first 3D device at 24 layers in 2013, with 160-layer parts in development by 2020. Capacity also comes from storing more bits per cell — one at SLC, two at MLC, three at TLC, four at QLC — and every step costs endurance, because more bits means finer voltage distinctions with less margin for wear.

That trade is most of what separates a server drive from a consumer one. Enterprise parts lean toward MLC and SLC, carry supercapacitors or batteries so in-flight writes survive a power cut, and come in U.2, U.3 and EDSFF form factors rather than the M.2 stick in a laptop. At the top of the range, Solidigm's D5-P5336 states up to 122.88 terabytes in one drive at up to 1,005,000 random read IOPS — against 36TB for the largest hard drive shipping commercially in 2025.

The interface had to change too, and here the numbers are stark. AHCI, designed for a spinning disk with one head that can only be in one place, allows a single command queue of at most 32 commands. NVMe allows up to 65,535 queues of up to 65,536 commands each. SATA 3.0 tops out at 6.0 gigabits per second; a PCIe 3.0 x4 link carries up to 31.5. Storage moved off a dedicated bus and onto the same general-purpose interconnect that carries accelerators and network cards.

None of which means the disk is dead, and the collection carries both halves of that. SSDs cost four to nine times more per bit; enterprise hard drives were around $14.4 per terabyte in 2022. Hard drive revenue and unit shipments are declining — and hard drives still hold most of the exabytes produced for servers. Flash takes what needs latency, disk keeps what needs cost per bit, and an AI training corpus measured in petabytes needs cost per bit.

Finally, why any of this belongs in a collection about AI. NVIDIA's own documentation for GPUDirect Storage says it exists to give a direct path between GPU memory and storage that 'avoids a bounce buffer through the CPU', because that extra copy 'introduces latency and lowers effective bandwidth' — and claims a direct path offers at least twice the peak bandwidth. The sentence underneath is the one that matters: the GPU, not the CPU, now has the first and last touch of data moving to and from storage. Storage stopped being a peripheral of the processor and became a peer of the accelerator.

What is missing is named rather than implied. JEDEC's JESD218B.01, the standard that actually defines how the endurance ratings quoted here are measured, sits behind a registration wall and was not read, so this page describes symptoms of endurance rather than its definitions. Solidigm's page did not state the endurance rating of its own flagship drive. Enterprise SSD pricing is not sourced anywhere here. And capacity claims for 245TB-class drives were set aside because they describe parts that have been demonstrated rather than shipped.

Figures

Every number below is asserted and sourced elsewhere on this page.

The density that stopped coming free

Annual growth in hard drive areal density by era. The final column is a manufacturer's forecast for a period that has not finished, not a measurement.

Areal density growth, per cent per year

60%/yr

1988–1996

100%/yr

1996–2003

30%/yr

2003–2010

10%/yr

2010–2016

20%/yr

2020–2034 (forecast)

Wikipedia's Hard disk drive article, read directly; the tertiary account of industry-wide rates no single manufacturer publishes neutrally. The first four columns are historical rates; the last is Seagate's forecast for 2020–2034 and had not happened when published.

More bits per cell, less life per cell

Bits stored in a single NAND cell by generation. Each step lowers cost per bit and lowers endurance, which is most of what separates a data-centre drive from a laptop one.

Bits stored per NAND cell

1 bit

SLC

2 bits

MLC (1998)

3 bits

TLC (2010)

4 bits

QLC (2009)

Wikipedia's Flash memory article, read directly. Bits per cell are definitional rather than measured; the article states that endurance decreases as bits per cell rise but gives no paired program/erase cycle figures, so none are shown.

What a data-centre drive holds

Single-drive capacity at the top of each technology. The flash part holds roughly three times the largest shipping disk — at four to nine times the cost per bit.

Total per drive, TB

  • Solidigm D5-P5336 (QLC flash)122.9 TB
  • Seagate Mozaic (HAMR disk)44 TB
  • Largest conventional disk, 202536 TB

Solidigm's product page for the D5-P5336 (122.88TB) and Wikipedia's Hard disk drive article (36TB largest shipping commercially as of 2025; Seagate Mozaic at 44TB from Seagate's own platform page). Shipping products, not roadmap figures.

Concepts

The vocabulary this subject is built from, and what we can show about each.

Areal Density and the Recording Roadmap

process

Hard drive areal density grew about 60% per year in 1988-1996 and 100% in 1996-2003, then fell to 30% in 2003-2010 and 10% in 2010-2016; Seagate forecasts about 20% per year for 2020-2034 — a recovery on the preceding decade, and under a third of the late-1990s rate.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Seagate's Mozaic platform states 4TB or more per disk and drive capacities up to 44TB from a ten-disk architecture, built on superlattice platinum-alloy media, a plasmonic writer with integrated nanophotonic laser, a spintronic reader and a 7nm controller, with a stated roadmap to 100TB.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read
  • SupportsPrimary evidenceRetrieved & read
    Mozaic Platform

    Seagate TechnologyCompany

The recording roadmap runs PMR from 2005, SMR from Seagate in 2013 as a stopgap, helium fill, energy-assisted recording in Western Digital's Ultrastar HC550 in late 2020, and HAMR shipping commercially in early 2024 after more than a decade of delay and a redesign of media, heads and lasers.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Disk Versus Flash Economics

other

SSDs cost four to nine times more per bit than hard drives, and enterprise drives were around $14.4 per terabyte in 2022; hard drive revenue and unit shipments are declining while hard drives still dominate the exabytes of storage produced for servers.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

DRAM (Dynamic Random-Access Memory)

memory technologyshared from another collection — see its own page for what it asserts

Enterprise Versus Client Storage

component

Solidigm states the D5-P5336 at up to 122.88TB in E3.S, E1.L and U.2 form factors and up to 1,005K 4KB random read IOPS for AI data lakes — several times the 36TB largest commercial hard drive of 2025, at several times the cost per bit.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

Enterprise drives differ from consumer ones in cell type (MLC/SLC against TLC/QLC), power-loss protection by supercapacitor or battery, unpowered retention after wear-out of about three months against a year, and server form factors U.2, U.3 and EDSFF rather than M.2 — against a consumer reference point of 600 TBW for a 2018 1TB drive.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Graphics Processing Unit (GPU)

componentshared from another collection — see its own page for what it asserts

High Bandwidth Memory (HBM)

memory technologyshared from another collection — see its own page for what it asserts

NAND Cell Levels

memory technology

NAND stores one bit per cell at SLC, two at MLC (NEC, 1998), three at TLC (Samsung mass production from 2010) and four at QLC (Toshiba and SanDisk, 2009), and endurance decreases as bits per cell rise — the trade-off behind most consumer-versus-enterprise drive differences.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Planar NAND halted around 16 nm and the industry stacked instead: Samsung commercialised the first 3D NAND at 24 layers in 2013 and had 160-layer parts in development by 2020 — the same move to a third dimension seen in stacked memory and multi-die accelerators.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

NAND Flash

memory technologyshared from another collection — see its own page for what it asserts

NVMe and the Queue

interconnect

AHCI allows one queue of at most 32 commands; NVMe allows up to 65,535 queues of up to 65,536 commands each, cuts uncacheable register accesses from nine per command to two, supports 2,048 MSI-X interrupts, and removes submission locking — a redesign for a medium with no head to move.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

SATA 3.0 tops out at 6.0 Gbit/s while a PCIe 3.0 x4 link carries up to 31.5 Gbit/s, moving storage off a dedicated bus onto the general-purpose interconnect that also carries accelerators and network adapters.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Storage in the AI Pipeline

other

NVIDIA's GPUDirect Storage provides a direct DMA path between GPU memory and storage to avoid a bounce buffer through CPU memory, which NVIDIA states introduces latency and lowers effective bandwidth; it claims a direct path offers at least twice the peak bandwidth of one routed through the CPU.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Timeline

What actually happened, in order, with sources.

Coverage
  • 4 United States
  • 1 South Korea
  • 1 Unattributed

Where this topic’s events took place, as far as our sources establish it. Events with no single location — a standards publication, say — and events we have not yet attributed are both counted as unattributed rather than omitted.

  1. 2024

    Heat-Assisted Magnetic Recording Finally Ships

    production milestoneUnited States

    Heat-assisted magnetic recording shipped commercially in early 2024 after more than a decade of delay, requiring redesigned media, heads and new lasers; Seagate's platform on it states 4TB or more per disk and up to 44TB per drive.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  2. 2024

    A Single Drive Passes 120 Terabytes

    production milestoneUnited States

    Solidigm states the D5-P5336 at up to 122.88TB per drive in E3.S, E1.L and U.2 form factors and up to 1,005K 4KB random read IOPS, positioned for AI data lakes — against 36TB for the largest commercial hard drive of 2025.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  3. 2020

    Western Digital Ships the First Energy-Assisted Drive

    production milestoneUnited States

    Western Digital shipped the Ultrastar HC550 in late 2020, the first energy-assisted magnetic recording drive, more than three years before heat-assisted recording reached commercial shipment.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  4. 2013

    Samsung Ships the First 3D NAND at 24 Layers

    production milestoneSouth Korea

    Samsung commercialised the first 3D NAND at 24 layers in 2013 after planar NAND halted around 16 nm, restoring cost-per-bit scaling by stacking rather than shrinking; Seagate introduced shingled magnetic recording the same year.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  5. 2005

    NAND Overtakes NOR, and Perpendicular Recording Ships

    technology generation milestone

    In 2005 NAND flash overtook NOR flash in sales, settling which flash architecture would scale, and perpendicular magnetic recording first shipped in hard drives.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  6. Date unknownno source establishes an occurrence date

    The Accelerator Starts Reading Storage Without the CPU

    technology generation milestoneUnited States

    NVIDIA's GPUDirect Storage gives the accelerator a direct DMA path to storage, avoiding a bounce buffer through CPU memory, with NVIDIA claiming at least twice the peak bandwidth of a CPU-routed path — no introduction date is established by the source read here.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read

Source register

All 21 sources behind this page — what we read, what we tried to read and could not, and what we looked at and set aside, with the reason in view for each. A concept shared with another collection brings its own references with it, so some entries here were surfaced for a neighbouring topic rather than this one.

Cited as evidence
7
Tried, could not read
0
Surfaced, set aside
14
Cited sources 7 distinct links

Original publisher links. Files open on the publisher’s site; we do not host copies. A linked document is not an additional source or an independent verification.

Surfaced, set aside14

These came up while researching and were deliberately not used. We do not claim to have read them — each is listed with why it was passed over, so the shape of the survey is visible and not just its conclusions.

Coverage & limits

What this page does and does not claim.

Sixth packet, and the third leg of the compute stack alongside memory and accelerators. Sixteen sources were consulted: seven were retrieved and read, and nine were surfaced and deliberately set aside with a stated reason each. Three of the seven are manufacturer primaries — Seagate's Mozaic platform page read in a browser, Solidigm's product page, and NVIDIA's GPUDirect Storage documentation, which is engineering documentation rather than marketing and is the strongest source here. The remaining four are tertiary encyclopedia articles, used for the industry-wide figures no single manufacturer publishes neutrally: areal density growth rates by era, the recording technology timeline, NAND cell levels, the AHCI-versus-NVMe queue comparison, and the enterprise-versus-consumer distinctions that vendors state inconsistently. Vendor superlatives are recorded as claims and never asserted; capacity claims for 245TB-class drives were set aside specifically because those parts have been demonstrated rather than shipped. Named gaps: JEDEC's JESD218B.01, which defines how the endurance ratings quoted throughout are actually measured, is behind a registration wall and unread, so this collection describes symptoms of endurance rather than its definitions; Solidigm's own page did not populate the endurance rating for its flagship drive; and enterprise SSD pricing is unsourced, the only figure available being an aggregator's explicit guess. Not yet editor-reviewed; every assertion reads as reported.

Source check, 2026-09-17. Numeric-presence checks passed for 17 assertions using available source text, which may be cached. This is not verification of their meaning. What this check does and does not prove →

  • Not editor-reviewed unless labelled. Assertions marked Reported are assembled from the sources shown and have not yet been checked by an editor. Only Primary source and Corroborated mean a human verified them.
  • Disagreements are preserved, not resolved. Where sources conflict, both accounts appear and the assertion is marked Disputed.
  • Retrieval status is disclosed per source. A source we could not open is never counted as evidence for an assertion.

This page is also available as structured data: /api/v1/topics/storage

Storage: Disks, Flash, and the Data an AI Model Eats — Manufacturing.ai | Manufacturing.ai