Plutux

Technology

NAND flash and SSDs: the tier that feeds and checkpoints the cluster

The storage under an AI cluster is flash, because nothing else delivers the aggregate bandwidth thousands of readers need. NAND is made by stacking memory cells vertically — hundreds of layers now — and turned into drives by a partly overlapping and partly separate set of companies.

In one sentence

NAND flash is non-volatile memory built as vertically stacked cell layers on a silicon wafer; solid-state drives package that flash with a controller and firmware into a storage device.

Flash stopped scaling sideways more than a decade ago and went vertical instead: cells are stacked in layers, and capacity grows by adding layers rather than by shrinking features. Layer counts in the hundreds are now normal, which turns manufacturing into an extreme etch problem — a single hole punched through the whole stack, straight, with no taper.

A drive is more than its flash. The controller manages wear, remaps failed blocks, handles the fact that flash must be erased in large units before it can be rewritten, and determines the drive's real-world behaviour under sustained load. Two drives with identical flash can perform very differently, and for AI workloads the sustained write behaviour during checkpointing is what matters.

How it works

Bits per cell trades endurance for cost

Storing more voltage levels in each cell packs more bits into the same silicon and makes each level harder to distinguish, which costs endurance and retention. The industry offers several points on that curve, and AI storage typically sits toward the higher-endurance end for write-heavy checkpoint workloads and the cheaper end for read-mostly datasets.

Why checkpointing stresses drives specifically

Every node writes its state simultaneously, in bulk, repeatedly through a run. That is a sustained sequential write burst rather than the mixed workload most drives are tuned for, and it exposes the difference between a drive's peak specification and what it sustains once its internal buffers are full.

A different supplier set from DRAM, mostly

Two of the three DRAM makers also produce NAND, joined by others with no DRAM business at all, and drive assembly extends to companies that buy flash rather than make it. That partial overlap is why NAND and DRAM cycles are correlated without moving together.

Western Digital is no longer one of them

Until February 2025 one company held both a hard-drive business and, through a long-running joint venture, a NAND business. They were separated: SanDisk took the flash operation and the joint venture, and Western Digital kept the drives. A supplier list that still shows the pre-separation company as a flash maker is describing a structure that no longer exists.

What this depends on

1 of these is marked as a chokepoint: a handful of qualified suppliers, a multi-year lead time, or a single geography.

  • Supply chainChokepoint

    High-aspect-ratio etch

    Punching a straight hole through hundreds of stacked layers is the defining process challenge of vertical flash.

    Plasma etch
  • Technology

    Controller silicon and firmware

    The controller determines wear levelling, error correction and sustained performance; the same flash behaves very differently behind different controllers.

    Semiconductor IP
  • Supply chain

    Silicon wafers

    Flash is built on the same polished substrates as logic, and a stacked device consumes a lot of them per usable terabyte.

    Silicon wafers
  • Standard

    Drive interface and form-factor specifications

    A drive is only a drop-in replacement because its command set, connector and dimensions are specified in common. Without that every fleet would be locked to the vendor it was built with.

What depends on this

Other pages in this map that name NAND and SSDs as something they cannot do without.

Who supplies this

What each company supplies at this step, and — where a public figure exists — its share of this specific market — with what that share measures, the period it covers and who published it. Some rows also show the company’s own reported revenue for the segment covering this step, which is a different thing: it says how much this business matters to that company, not how much of the market it holds. Not a ranking and not a recommendation.

  • Samsung Electronics005930.KS· Korea

    The largest NAND manufacturer, and the only supplier that also makes the controllers and the drives.

  • Manufactures NAND and owns Solidigm, the enterprise drive business bought from Intel.

  • Manufactures NAND flash and drives; the business Toshiba's memory operation became.

  • SanDiskSNDK

    Manufactures NAND through its long-standing joint venture with Kioxia, and sells drives and retail flash.

  • Micron TechnologyMU

    Manufactures NAND flash and drives alongside its DRAM business.

    NAND Products segment — 22.9% of Micron Technology’s reported revenue ($8.5B)FY2025 · company filings · not a market share

  • Yangtze Memory TechnologiesPrivate

    Chinese NAND manufacturer, state-backed and under export restrictions.

  • Phison ElectronicsTaiwan

    Supplies the controllers and firmware most drives that are not Samsung's are built around.

  • Silicon MotionSIMO

    The other merchant controller house, supplying the drive makers who do not design their own silicon.

  • Marvell TechnologyMRVL

    Supplies the enterprise SSD controllers inside the drives data-centre buyers qualify.

  • Lam ResearchLRCX

    Supplies the high-aspect-ratio etch that cuts through the stack; layer count rises only as fast as this tool allows.

  • Tokyo Electron8035.T· Japan

    Supplies the deposition and coating steps that lay down each layer before it is etched.

  • Kingston TechnologyPrivate

    The largest independent assembler of drives and modules, buying flash from every manufacturer above.

What would change the picture

  • Whether layer counts keep rising at the pace that has held cost per bit falling.

  • Whether AI checkpoint workloads shift purchasing toward higher-endurance drives.

  • Whether flash displaces spinning disk in the cold tier as the cost gap narrows.

Questions people ask about this

Why did flash go vertical?
Because shrinking cells sideways stopped working: cells became too close to hold charge reliably and too small to distinguish voltage levels. Stacking layers grows capacity without shrinking the cell, which trades a two-dimensional scaling problem for an extremely demanding etch problem.
Does the drive matter or just the flash?
Both, and for AI the drive often matters more. The controller decides how the drive behaves under sustained writes, which is exactly what checkpointing produces. Peak specifications are measured on short bursts; what a cluster experiences is the sustained figure after internal buffers fill.

How these pages are written

Each page explains one technology in plain language, states what it depends on, and names companies by what they supply at that step. Company roles are described qualitatively and deliberately carry no market shares, revenue figures or rankings — those change faster than an explainer can, and a stale number is worse than none. Ticker links point at company pages on this site and are provided for reference only.

Nothing here is investment advice, a recommendation, or a forecast. A company named on a page about a technology is not thereby a good investment, and the chokepoints described are structural facts about supply chains rather than predictions about prices. Technology moves; where a page describes something as unresolved or in development, that was true when it was written.

Plutux is not an investment adviser. Market data and AI-generated analysis are for information and education only, not investment advice. Disclaimer

© Plutux Technology Limited 2026