Issue 10 · Week of Oct 4–11, 2026Weekly Storage & memory infrastructure for AI

The Storage and Memory Signal

Weekly signal for the people building and buying AI infrastructure: what moved, why it matters, what's next.

By AiInfraStorage.com Editors

← Previous issue

Micron spent the week confirming a thesis this newsletter has tracked since Issue 03: once HBM eats enough DRAM-capable wafer capacity, the company making it gets to set next year's price before next year starts. Fiscal Q4 revenue hit a record $54.23 billion, and on the earnings call Micron's own COO said calendar 2027 HBM pricing has already been reset higher, with more than three-quarters of that year's output already spoken for. TrendForce backed the pricing story from a different angle, forecasting Q4 contract prices up 10 to 15 percent for DRAM and 15 to 20 percent for NAND, even as the holiday-quiet spot market moved the other way. NetApp used its INSIGHT conference to launch a disaggregated file system claiming 100 terabytes a second and a new storage partnership with Oracle, and Everpure wasted no time answering with its own AI-context tools. SNIA's Developer Conference also wrapped in Santa Clara this week, though post-event coverage has been thin enough that we're saying so plainly rather than pretending otherwise.

01: The pulse

This week in four numbers

$54.23B Micron's record fiscal Q4 2026 revenue, up from $11.32 billion a year earlier as AI-driven DRAM and NAND pricing both spiked
15-20% TrendForce's forecast jump in NAND contract prices quarter over quarter for Q4 2026, with DRAM contract prices seen rising a further 10-15%
100TB/s Claimed throughput of NetApp's new Novus file system, built to let AI-factory metadata and data nodes scale independently
4.8x Per-GPU inference token throughput CoreWeave claims after putting Nvidia's Vera Rubin NVL72 into production, versus its GB200 NVL72 baseline
02: Shortage watch

The standing read on memory and NAND supply

A persistent line on where the shortage stands, present every issue whether or not the picture actually moved.

Direction: Worsening.

This period: On Micron's fiscal Q4 2026 earnings call September 30, President and COO Manish Bhatia said the company has "increased that pricing significantly for calendar year '27, which will reset at the beginning of the calendar year to narrow the profitability gap with conventional DRAM," and that more than three-quarters of fiscal 2027 HBM output is already committed. That same day, TrendForce forecast conventional DRAM contract prices rising 10 to 15 percent and NAND contract prices 15 to 20 percent quarter over quarter in Q4 2026, citing continued AI infrastructure procurement from cloud providers. Working against that read in the short term, TrendForce's own spot-price tracker showed a quiet holiday week: DDR4 1Gx8 spot roughly flat and the 512Gb TLC NAND wafer spot price down 2.45% week over week. That's soft consumer-segment spot trading, though, not the enterprise contract pricing and HBM allocation this call actually tracks.

Driver: HBM production is still eating disproportionately into wafer capacity that would otherwise make conventional DRAM and NAND, the same driver this newsletter has tracked since Issue 03, now confirmed in a memory maker's own forward pricing language rather than only in analyst inventory estimates or capacity-shift reporting.

Watch for: Samsung's and SK hynix's own Q3 2026 earnings calls in late October, the first chance to see whether either follows Micron in explicitly resetting calendar 2027 HBM contract pricing higher.

03: The index

Tracked over time

A running index of a few figures we track issue over issue, so the trend is visible, not just the snapshot. Three fresh readings clear the bar this week, all off TrendForce: the 512Gb TLC NAND wafer spot price fell again in a quiet week, DDR4 spot ticked up slightly, and DDR5 16Gb spot inched higher on TrendForce's live pricing page, since TrendForce's own September 30 roundup didn't include a dedicated DDR5 figure the way it did for DDR4 and NAND. Nothing new on the 30TB SSD or 30TB HDD list prices this issue.

IssueDateMetricValue
Issue 03Aug 30, 202630TB enterprise SSD list price$22,600
Issue 04Sep 4, 202630TB enterprise SSD list price$22,600 (flat)
Issue 05Sep 1, 2026DDR4 1Gx8 3200MT/s spot price$44.54 (+2.08% WoW)
Issue 05Aug 31, 2026512Gb TLC NAND wafer spot price$20.71 (-0.90% WoW)
Issue 07Sep 11, 2026DDR4 1Gx8 3200MT/s spot price$45.21 (-0.14% WoW)
Issue 07Sep 11, 2026DDR5 16Gb (2Gx8) 4800/5600 spot price$54.33 (flat WoW)
Issue 08Sep 14, 2026512Gb TLC NAND wafer spot price$20.083 (-0.31% WoW)
Issue 08Sep 15, 2026DDR4 1Gx8 3200MT/s spot price$45.54 (+0.47% WoW)
Issue 09Sep 21, 2026512Gb TLC NAND wafer spot price$19.883 (-1.00% WoW)
Issue 09Sep 24, 2026DDR4 1Gx8 3200MT/s spot price$46.107 (+1.2% since Sep 15)
Issue 09Sep 26, 2026DDR5 16Gb (2Gx8) 4800/5600 spot price$57.667 (+6.1% since Sep 11)
Issue 10Sep 30, 2026512Gb TLC NAND wafer spot price$19.396 (-2.45% WoW)
Issue 10Sep 30, 2026DDR4 1Gx8 3200MT/s spot price$46.32 (+0.93% WoW)
Issue 10Oct 2, 2026DDR5 16Gb (2Gx8) 4800/5600 spot price$58.00 (+0.12% session)
04: Signal

What moved, and why it matters

Five things that happened this week, and why I'd pay attention to each one.

SNIA's Developer Conference wrapped in Santa Clara, with coverage still thin SNIA · GlobeNewswire

The SNIA Developer Conference ran September 28-30 at the Hyatt Regency Santa Clara, with the SMB3 Interoperability Lab continuing through October 1, exactly as this newsletter previewed last issue. The pre-event agenda promised a StorageAI track covering KV-cache storage offload, with sessions from Dell and Cerebras, and vector-database indexing offload from Solidigm, plus a keynote from Dell's Jason Duquette and a 10th-anniversary celebration for SNIA's Swordfish management standard. As of this issue's publication, SNIA's own newsroom and blog hadn't posted anything dated after the event closed, and no vendor had put out a post-show release describing what was actually said or demonstrated on the StorageAI stage.

Worth notingThis newsletter committed to covering SDC at Signal level at minimum once it wrapped. The honest version of that this week is: here's what was scheduled, and the write-ups haven't shown up yet. We'll follow up the moment SNIA or a session vendor actually publishes what happened.
CoreWeave puts Nvidia's Vera Rubin NVL72 into production CoreWeave · StorageReview

CoreWeave said October 2 that AI coding startup Cognition is now running production inference on Nvidia's Vera Rubin NVL72 rack-scale system, the first live customer workload on the platform. CoreWeave's own benchmarks claim up to 4.8x more inference tokens per GPU than a GB200 NVL72 baseline. The new Vera CPU compute nodes behind the deployment pair dual 88-core Arm CPUs with 1.5TB of system memory and 15.36TB of local NVMe per node, a meaningfully bigger local-storage footprint than CoreWeave's prior GB200-generation nodes carried.

Worth watchingEvery generational jump in GPU rack density seems to come with a matching jump in local NVMe per node, the quiet storage story inside every flashy new accelerator launch. Whether a 4.8x inference claim holds up outside CoreWeave's own benchmark is the open question, same as every vendor-reported multiple this year.
Everpure answers NetApp and HPE with Pure KVA and Always-On DeepReduce Everpure · Blocks & Files

Everpure, the storage vendor formerly known as Pure Storage, used the same week as NetApp's INSIGHT announcements to roll out its own AI platform additions: Pure KVA, which pre-stages AI context data into GPU memory ahead of a request and claims up to 20x faster time-to-first-token, Always-On DeepReduce, a data-reduction engine the company says delivers a median additional 2:1 reduction on top of existing compression, and Everpure Data Intelligence, a cross-silo data discovery and classification layer that exposes data to AI agents through MCP, the Model Context Protocol.

Why it's hereThat's five storage vendors now pitching some version of pre-staging a KV cache into GPU memory before the model asks for it: NetApp, HPE, Huawei, Silicon Motion, Kioxia, and now Everpure. This issue's Second Cut goes deeper on NetApp's own entry, but Everpure's 20x time-to-first-token claim is the same specific multiple HPE used in Issue 08 for an identical pitch, which should make any buyer want a shared, independent benchmark rather than six vendors each grading their own homework.
TrendForce: Q4 contract prices are about to jump again, even as spot cools TrendForce

TrendForce forecast September 30 that conventional DRAM contract prices will climb 10 to 15 percent quarter over quarter in Q4 2026 and NAND flash contract prices 15 to 20 percent, pointing to cloud providers continuing to expand AI infrastructure procurement as the driver. That call landed the same week TrendForce's own spot-price tracker moved the other way for the short term: DDR4 1Gx8 spot roughly flat and the 512Gb TLC NAND wafer spot price down 2.45 percent week over week, in what the company described as a quiet holiday week with weak NAND buying.

Worth notingContract price is what shows up on an actual hyperscaler purchase order; spot price is a much smaller, noisier side market. A forecast contract jump alongside a soft spot week isn't a contradiction, it's a reminder that these are two different markets, the exact distinction Issue 06's Uplevel Highlight on spot versus contract pricing was built to explain.
SK hynix, again, says no decision has been made on a Solidigm IPO Korea Times · Digitimes

SK hynix repeated on October 1 that it is "reviewing various funding options" for Solidigm but has made no decision on a US listing, pushing back on the $100 billion to $150 billion valuation range Reuters and Bloomberg reported in last issue's Deep Cut. It's the same holding language the company has used since August, now in its fourth straight month of circulation, with Solidigm separately still said to be studying a US NAND fab site, upstate New York among the candidates.

Worth watchingSK hynix gave itself until on or around November 5 to detail Solidigm's strategic review, a deadline this newsletter has now tracked across four issues without it slipping. Another month of "reviewing options" language means that date, not this week's non-news, is still the one that matters.
05: Vendor tracker

Where the big three stand on HBM

A running snapshot, updated whenever a vendor discloses something new, not re-explained from scratch every week. Micron disclosed a significant forward-pricing and commitment update on its earnings call this issue; Samsung and SK hynix carry forward unchanged.

SamsungUpdated Issue 09

CUBE roadmap (Issue 08) still stands: HBM5 targets 2x HBM4E performance, zHBM benchmarked at 8x HBM4E, zNAND-O sampling penciled in for 2028. Industry sourcing from Issue 09 says Samsung will raise monthly HBM wafer input from roughly 180,000 to roughly 250,000 next year and lift HBM4/HBM4E's share of output from about 40% to about 80%.

HBM wafer input rising ~180K to ~250K/month, 2026 to 2027 plan Capacity ramp
SK hynixUpdated Issue 03

Hybrid bonding pushed from HBM4E to HBM5 after hitting a 775-micron packaging ceiling. Racing for 16-Hi HBM4 delivery to Nvidia by Q4 2026.

Delayed
MicronUpdated Issue 10

Reportedly adding up to 60K HBM wafers per month toward roughly 100K by year end (Issue 05); still racing for 16-Hi HBM4 by Q4 2026. New this issue: fiscal Q4 2026 earnings call disclosed that calendar 2027 HBM pricing has been reset significantly higher to narrow the profitability gap with conventional DRAM, with more than three-quarters of fiscal 2027 HBM output already committed. CTO Scott DeBoer said Micron has worked with Nvidia for over a year on a custom HBM4E product for next-generation GPU and NVLink Fusion platforms. A second HBM packaging facility in Taichung, Taiwan is set to start up this month.

More than 75% of fiscal 2027 HBM output already committed, per Q4 FY26 earnings call Pricing reset
06: Storage leaderboard

Where the enterprise storage systems vendors rank

No new IDC or Gartner report landed this issue, so every card below still reflects IDC's Q2 2026 Worldwide Quarterly Enterprise Storage Systems Tracker and the 2026 Gartner Magic Quadrant, both covered in Issue 08.

DellIDC Rank 1 · Gartner Leader

IDC's Q2 2026 tracker put Dell's external enterprise storage revenue at $2.46 billion, a 23.8% share and the fastest year-over-year growth (+42.5%) among the top vendors, riding an AI-storage attach strategy across PowerScale and PowerStore.

Last updated Issue 08
HuaweiIDC Rank 2 · Gartner Leader

Jumped from fourth to second in IDC's Q2 2026 tracker at an 11.3% share ($1.17 billion), passing NetApp; still strongest outside North America.

Last updated Issue 08
NetAppIDC Rank 3 · Gartner Leader

Slipped to third in IDC's Q2 2026 tracker (9.6% share, $988 million), passed by Huawei; AFX disaggregated architecture and the AI Data Engine (AIDE) remain its core pitch.

Last updated Issue 08
EverpureIDC Rank 3 (Q1) · Gartner Leader

Formerly Pure Storage, renamed February 2026. Number one on both Magic Quadrant axes for the second straight year; IDC's Q2 tracker didn't name its exact share, so this reflects Q1's ranking.

Last updated Issue 03
HPEIDC Rank 5 (Q1) · Gartner Leader

Folding storage into its broader AI factory and GreenLake positioning.

Last updated Issue 01
IBMGartner Leader

IBM Storage Scale holds roughly 17% of the parallel file system market specifically; Storage Scale System 6000 is NVIDIA-certified for metadata-heavy training jobs.

Last updated Issue 01
07: Deep cut

Micron already sold 2027

A record $54 billion quarter came with a bigger disclosure buried in the call: most of next year's HBM output is already spoken for, at a price Micron reset itself.

Micron reported fiscal fourth-quarter revenue of $54.23 billion on September 30, a record, up from $11.32 billion in the same quarter a year earlier. That 379% year-over-year jump looks like a typo until you remember how weak the comparison quarter was, before this year's AI memory ramp properly began; Micron states the same figure itself in its own release. Full fiscal 2026 revenue came to $133.19 billion.

By product type, Micron said DRAM revenue reached $39.8 billion, 73% of the total and up 343% year over year, with average selling prices up in the high teens percent sequentially. NAND brought in $14.10 billion, with ASPs up roughly 30% sequentially; data center SSD revenue alone came to around $10 billion, more than ten times what it was a year ago and over two-thirds of total NAND revenue. Micron's own regulatory filing cuts the same quarter a different way, by business unit rather than product: Cloud Memory at $16.28 billion, Core Data Center at $18.00 billion, Mobile and Client at $13.11 billion, and Automotive and Embedded at $6.82 billion. Both views describe the same $54 billion quarter; neither breaks out HBM revenue specifically, still true across the whole industry, not just Micron.

FQ4 2026 revenue$54.23B, record, vs $11.32B a year ago; full FY2026 revenue $133.19B
DRAM / NAND splitDRAM $39.8B (73% of revenue, +343% YoY); NAND $14.10B, data center SSD ~$10B (>10x YoY)
Business-unit cut (8-K)Cloud Memory $16.28B; Core Data Center $18.00B; Mobile & Client $13.11B; Automotive & Embedded $6.82B
HBM disclosureCOO Manish Bhatia: CY2027 HBM pricing reset significantly higher; >75% of FY2027 HBM output already committed
FQ1 FY2027 guidanceRevenue $61.5B ± $1.5B

What Micron disclosed on the call, rather than in the press release, matters more than any of those totals. President and COO Manish Bhatia told analysts the company has "increased that pricing significantly for calendar year '27, which will reset at the beginning of the calendar year to narrow the profitability gap with conventional DRAM," and said more than three-quarters of fiscal 2027 HBM output is already committed. That's Micron, not an analyst or a channel check, confirming in its own words that this isn't simply a tight market working itself out on its own. It's a supplier that has already sold most of next year's most profitable product, at a price it set itself, three months before that year starts. CTO Scott DeBoer added that Micron has been working with Nvidia for more than a year on a custom HBM4E product for next-generation GPU and NVLink Fusion platforms, the NV-HBM concept this newsletter has flagged in earlier roadmap coverage.

Micron is backing that commitment with physical capacity, too. Digitimes reported October 2 that Micron is expanding HBM-related capacity across four Taiwan sites, Taoyuan, Taichung, Tongluo, and Tainan, with a second HBM packaging facility in Taichung due to start up this month. Separately, Applied Materials said it is opening a Silicon Valley research hub, the EPIC Center, dedicated to memory and advanced-packaging work for the AI era, with Micron, Samsung, SK hynix, TSMC, and Broadcom already members and Kioxia and Besi confirmed as the newest to join. Direct rivals sharing a packaging research table is itself a signal: the bottleneck has moved from who can etch the smallest DRAM cell to who can stack and bond dies fast enough to keep up, a problem big enough that competitors would rather pool some basic research than solve it alone.

Read the whole quarter next to CEO Sanjay Mehrotra's own framing, that "AI is becoming Super Intelligence (SI), and memory enhances this intelligence and competitiveness of our customers' platforms." Strip the branding and the mechanics underneath are less mystical: a supplier facing effectively unlimited demand for a product it can't make fast enough gets to name its price for a year that hasn't started, and buyers sign anyway because the alternative is no supply at all. That dynamic, confirmed now in Micron's own call rather than inferred from inventory data, is the clearest sign yet that Issue 03's thesis about HBM eating conventional DRAM capacity hasn't just continued through Q3. It's hardened into Micron's actual pricing policy for next year.

Worth watchingWhether Samsung and SK hynix make the same kind of explicit 2027 pricing disclosure on their own Q3 calls in late October, or whether Micron is the only one willing to say the quiet part about HBM pricing out loud.
08: Second cut

NetApp's double header: a 100TB/sec file system and a seat inside Oracle's cloud

Two announcements in two days, aimed at two different problems: NetApp's slipping rank in IDC's own tracker, and a storage architecture AI factories are outgrowing.

NetApp used its INSIGHT conference in Las Vegas, running September 29 through October 1, to launch Novus, a new file system the company says disaggregates metadata and control-plane servers from the data and capacity nodes underneath, letting each scale independently inside a single NFS namespace. NetApp's own claim is more than 100 terabytes per second of aggregate throughput and support for clusters running "hundreds of thousands of GPUs." Novus ships alongside Novus Data Director and sits on existing ONTAP and AFF A90 hardware, and it's orderable now, not a concept demo.

The rest of INSIGHT's announcement list filled out a broader pitch: Keystone Sovereign, targeting EU data-sovereignty requirements with pilots in Germany and France; Console Fleet Management; an expansion of the AI Data Engine (AIDE); a ChatOps-style administrative interface; and new partnership language with SAP, Commvault, and Supermicro. A day earlier, on September 29, NetApp and Oracle separately announced a fully managed cloud storage service bringing ONTAP natively into Oracle Cloud Infrastructure for AI and enterprise workloads, giving NetApp another hyperscaler storefront to sell through beyond its existing AWS, Azure, and Google Cloud deals.

NovusDisaggregated metadata/data file system, claimed >100TB/sec throughput, built for "hundreds of thousands of GPUs," orderable now
Also at INSIGHTKeystone Sovereign (EU pilots, Germany/France); Console Fleet Management; AIDE expansion; ChatOps admin interface; new SAP/Commvault/Supermicro partnerships
NetApp + OracleFully managed ONTAP-native storage service for OCI, announced Sept 29
Context, Issue 08NetApp slipped to third in IDC's Q2 2026 enterprise storage tracker, passed by Huawei
Also this monthNetApp announced intent to acquire PEAK:AIO (Sept 25), adding metadata services and parallel-namespace technology

The timing isn't subtle. NetApp dropped to third place in IDC's own Q2 2026 enterprise storage tracker two issues ago, passed by Huawei, and a disaggregated architecture claiming the fastest file system numbers in the business is exactly the kind of headline a vendor needs to change that conversation before the next quarterly tracker drops. It's also worth noting that Novus's metadata-disaggregation pitch lands just six days after NetApp said it would acquire PEAK:AIO specifically for metadata services and parallel-namespace technology. NetApp hasn't said the two are connected, and six days is a tight window to have actually integrated anything, but it's a detail worth watching as Novus ships more broadly.

The 100TB/sec figure is NetApp's own number, with no independent benchmark attached yet. This newsletter has tracked enough vendor-reported multiples this year, HPE's 20x in Issue 08, Everpure's own 20x this issue, to treat any single-vendor throughput claim as a marketing opening bid rather than a settled fact. The next real test is whether Novus shows up in a future MLPerf Storage submission round, the one place these numbers get checked by someone other than the company making them.

Why it's hereThis issue's Uplevel Highlight goes deeper on what "disaggregated" actually means in a storage architecture, and why AI training clusters are the workload finally forcing vendors to build it for real.
09: Uplevel highlight

This week's term: Disaggregated storage

One piece of storage or memory vocabulary, explained properly, every week.

Disaggregated storage

In plain English: splitting a storage system into two separate pools, one that tracks where everything is, and one that actually holds the data, so each can be scaled up on its own instead of being stuck in a fixed ratio to each other.

A traditional storage array ships as a bundle. The controllers that track file locations, permissions, and namespace structure sit in a fixed pairing with the drives that actually hold the bytes, and buying more of one usually means buying more of the other, whether a given workload needs it or not. That bundling was a reasonable trade-off for decades of general-purpose enterprise storage, where metadata load and raw capacity tended to grow at roughly similar rates.

AI training and inference broke that assumption. Checkpointing a large model writes an enormous burst of data all at once, stressing the capacity side. A cluster running thousands of GPU workers against a shared namespace, or an inference fleet constantly reading and writing KV cache, stresses the metadata side instead, often independently of how much raw capacity is actually in use. A system where both layers are welded together gets bottlenecked on whichever one is under more load, even when the other has headroom to spare. Disaggregation separates the piece that answers "where is this file and who can touch it" from the piece that moves the actual bytes, letting an operator add more of whichever side is actually under strain.

NetApp's Novus, covered in this issue's Second Cut, is this week's concrete example, but the idea isn't new or NetApp's alone. VAST Data built its DASE architecture around the same separation years ago, and WEKA's NeuralMesh does a version of it as well. A buyer evaluating this kind of claim has several products to compare, not just one vendor's marketing term for a concept the whole high-performance storage category already uses in some form.

The honest caveat: "disaggregated" has become a label vendors reach for whenever they want to signal they've modernized, and the throughput multiple attached to any one implementation deserves the same skepticism this newsletter applies to every other vendor-reported number. A useful test for telling real disaggregation from a relabeled old architecture: ask whether the metadata tier and the data tier can actually be bought, licensed, and scaled on separate schedules, or whether they still ship in a fixed bundle regardless of what the brochure calls it.

Why it's in this newsletterEvery AI training cluster eventually hits a wall that isn't about raw storage capacity, it's about how fast the system can answer "where is this" millions of times a second while also moving petabytes. Disaggregation is the storage layer's answer to that specific wall, not a synonym for "new."
10: On the radar

What's coming up

OCT 12–15

OCP Global Summit 2026, San Jose McEnery Convention Center. Nvidia's Ian Buck is confirmed to keynote "AI Factories at Scale" on October 12, with Lattice Semiconductor CEO Ford Tamer also keynoting October 13. The last Global Summit before it relocates to San Francisco's Moscone Center starting in 2027.

OCT 28–30

Seagate's next quarterly earnings call, scheduled October 28, and Kioxia's fiscal Q2 2026 results on October 30, the first test of whether HDD and NAND pricing are following the contract-price jump TrendForce is forecasting for the quarter.

LATE OCT

SK hynix's Q3 2026 earnings call, the next chance to see whether it follows Micron in explicitly resetting calendar 2027 HBM contract pricing higher.

~NOV 5

SK hynix's self-imposed window to detail Solidigm's strategic review runs out, four issues running now without a resolution.