Key points
- An AI cluster stores data in three tiers: HBM next to the GPU, NAND flash SSDs, and nearline hard drives.
- Sandisk (SNDK) and SK hynix published the first spec for High Bandwidth Flash in August 2026, a proposed new tier with several times HBM's capacity.
- Flash repriced about three times harder than hard drives over the past year: 84.6% gross margins at Sandisk and Micron (MU) against 52% to 54% at Seagate (STX) and WD.
Every byte an AI model trains on or serves back has to live somewhere, and where it lives is a price ladder. The fastest tier is high-bandwidth memory, the DRAM stacked next to the GPU that holds model weights while the model runs. Below that sits NAND flash, sold as enterprise solid-state drives, where training data gets staged and where inference reads from when something isn't already in memory. Underneath everything are nearline hard drives, the cheap high-capacity disks holding the archive a data center can't afford to keep on flash.
Four US-listed companies cover the ladder: Micron (MU) at the top, Sandisk (SNDK) in the middle, and Seagate (STX) and Western Digital (WDC) at the base. The margins tell part of the story, and what each one is actually building for the AI cycle tells the rest, because the roadmaps run further than any quarterly number.
What each company is building
Micron shares the HBM tier with SK Hynix and Samsung, the trio our HBM explainer covers. Its HBM4 part, first sampled to customers in June 2025 as a 36GB twelve-high stack, moves more than 2.0 terabytes per second per stack, and the company said the ramp is timed to customers' next-generation AI platforms in calendar 2026. Chairman, president and chief executive Sanjay Mehrotra said in the company's June 24, 2026 results that "Micron is investing at record levels in technology, products and supply to address our customers' rapidly growing demand." That quarter, reported June 24, put revenue at $41.46 billion against $9.30 billion a year earlier.
Sandisk's flash sits one rung down, and its bet is that the rung moves up. In August 2025 it began working with SK hynix on High Bandwidth Flash, NAND stacked and interfaced like HBM so a GPU can read from it directly. Sandisk has said HBF targets bandwidth in HBM's range with several times the capacity for similar cost, and the first technical specification, published through the Open Compute Project on August 3, 2026, describes stacks of up to 512GB with host bandwidth tiers running from 400 gigabytes per second to 3 terabytes per second. Sandisk has targeted first HBF samples in the second half of 2026, with sample AI-inference devices built on it in early 2027. If it works as specified, inference systems get a layer that holds far larger models near the GPU without paying HBM prices. Today's business is enterprise SSDs, which is the line that grew 437% in the fiscal year Sandisk reported on August 5, 2026, and the ownership history behind rival Kioxia is tangled up with Sandisk's own.
The drive makers are running a different race: capacity per disk. Seagate's HAMR-based Mozaic 4+ platform, announced March 3, 2026, ships drives up to 44TB and is qualified and in production with two hyperscale cloud providers, on a roadmap the company says scales from just over 4TB per disk toward 10TB per disk, which means hard drives of up to 100TB. Chair and chief executive Dave Mosley, in the company's July 29, 2026 results: "As AI accelerates data generation and its value, we see durable long-term demand for mass capacity storage."
WD is a step behind on HAMR and says so in its roadmap. At its February 3, 2026 innovation day it put a 40TB UltraSMR drive, built on its current ePMR recording, in qualification with two hyperscale customers ahead of volume production in the second half of 2026, with HAMR qualifications underway and ramp production planned for 2027. The company says ePMR extends to 60TB and HAMR scales to 100TB by 2029. Chief executive Irving Tan framed the year's work as having "enabled us to truly reimagine the hard drive to meet the requirements of AI."
| Company | What it's building for AI | Latest reported quarter |
|---|---|---|
| Micron Technology (MU) | HBM4: 36GB stacks moving 2.0+ TB/s, ramping calendar 2026 | Revenue $41.46B, gross margin 84.6% (FQ3, reported June 24, 2026) |
| Sandisk (SNDK) | Enterprise SSDs; High Bandwidth Flash spec published August 2026 | Revenue $8.97B, gross margin 84.6%, datacenter up 437% for the year (FQ4, reported August 5, 2026) |
| Seagate Technology (STX) | HAMR Mozaic 4+ at 44TB, roadmap to 100TB drives | Revenue $3.63B, gross margin 52.3% (FQ4, reported July 29, 2026) |
| Western Digital (WDC) | 40TB UltraSMR ramping H2 2026, HAMR ramp 2027, 100TB by 2029 | Revenue $3.75B, up 44%, gross margin 54.1% (FQ4, reported August 5, 2026) |
What the last year's numbers say
The financials all point the same direction and land very differently by tier. In the quarters reported between late June and early August 2026, Sandisk's gross margin reached 84.6% against 26.2% a year earlier, and Micron's reached 84.6% against 37.7%. Seagate went from 37.4% to 52.3% and WD from 41.0% to 54.1%. Same shortage, but flash roughly tripled its margin while hard drives added about thirteen points, because nearline drives never got as short as flash did.
Both flash makers spent the year converting spot demand into contracts. Sandisk said that since announcing five New Business Model agreements in April it signed five more, including three with new customers, and Micron titled its June release around executing "transformational Strategic Customer Agreements." Sandisk's board also authorized an additional $14 billion of buybacks, taking total remaining authorization to $15.5 billion against full-year net income of $11.43 billion. That's capital going back to shareholders rather than into new fabs, at a moment when the shortage would reward more supply.
What could go wrong
Sandisk itself published the thing to worry about. It said sequential revenue growth came "approximately one-third from higher volumes and two-thirds from higher pricing." Two-thirds of the increase is the same product sold dearer, and memory pricing has historically gone both directions hard. A margin that traveled 58 points up in a year can travel down, which is why the capability roadmaps matter more than any single quarter: they're what each company keeps if pricing turns.
The roadmaps carry their own risks. High Bandwidth Flash is a specification and a sampling target, not a shipping product, and it needs GPU makers to design for it. HAMR took Seagate well over a decade to bring to volume, and WD's version is still in qualification, so the 100TB numbers on both roadmaps are destinations rather than products. And all four companies sell to the same handful of hyperscalers funding everything else in the AI build, so storage demand isn't independent of the capex cycle it sits inside.
None of this makes any of them a buy or a sell. It maps which tier each company sells into, what each one is building next, and what each last reported. Sandisk chairman and chief executive David Goeckeler, on the fiscal year: "We closed fiscal 2026 with a leading technology portfolio, established datacenter as a key growth pillar, and deepened our customer partnerships."
Sources
- Sandisk, fiscal fourth quarter 2026 results (August 5, 2026) and the High Bandwidth Flash OCP specification announcement (August 3, 2026)
- Micron Technology, fiscal third quarter 2026 results (June 24, 2026) and the HBM4 customer sampling announcement
- Seagate Technology, fiscal fourth quarter 2026 results (July 29, 2026) and the Mozaic 4+ announcement (March 3, 2026)
- Western Digital, fiscal fourth quarter 2026 results (August 5, 2026) and the AI-era storage roadmap announcement (February 3, 2026)
- Related coverage: the AI hardware stack, layer by layer
This is general market commentary and opinion, not investment advice. Markets can go down as well as up, and you can lose money. Always do your own research and consider speaking with a licensed financial professional before making any investment decision.
