Kioxia GP1: 20M IOPS SSD Beats DRAM? [2026]

Kioxia Corporation is pushing its flash memory lineup into territory once reserved for DRAM, and the numbers behind that push are drawing attention across the storage industry this week. According to a report from TechRadar, an upgraded version of Kioxia’s GP1 Series PCIe 6.0 SSD now exceeds 20 million random-read IOPS, a figure that, if it holds up, would roughly double the performance Kioxia’s own official rollout promised just a few months earlier. The timing matters. AI data centers are running short on the kind of ultra-fast, low-latency memory that keeps GPUs fed, and flash vendors smell an opening.

This article breaks down what Kioxia has actually confirmed, what is still being reported rather than officially verified, and why a storage company is suddenly talking about challenging DRAM instead of just competing with other SSD makers.

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

What Kioxia Confirmed About the GP1 Upgrade

Kioxia’s own press release, dated August 4, 2026, introduced the KIOXIA GP1 Series as the first product in a new GP Series lineup the company describes as its first super-high-IOPS PCIe SSD built specifically for GPU direct access. The official figure Kioxia put on record was up to 10 million random-read IOPS, measured using a 512-byte block size, running over a PCIe 6.0 interface with NVMe 2.2 support. That is the number that appears in Kioxia’s own materials and in the press release distributed around the original launch.

The GP1 drive is built around KIOXIA XL-FLASH generation 2, a low-latency flash memory type Kioxia has been refining for several product cycles. The company positions the drive as a flash-based memory-extension tier, something that sits alongside High Bandwidth Memory (HBM) in an AI server rather than replacing it outright. That framing is important, because it tells you Kioxia isn’t claiming the GP1 is a DRAM swap today. It’s claiming the GP1 is step one on a roadmap toward that goal.

The 20 Million IOPS Claim: What’s Confirmed vs. What’s Reported

Here’s where readers need to slow down. The 20-million-IOPS figure, the “world record” framing, and the “fastest SSD ever produced” label all come from TechRadar’s reporting on an upgraded edition of the GP drive, not from Kioxia’s original, cited press material. Kioxia’s own GP1 announcement states up to 10 million random-read IOPS. TechRadar’s account describes a faster edition that exceeds 20 million IOPS under the same 512-byte random-read test condition, alongside a separate reference to a third-generation XL-FLASH design that improves memory access through shorter wiring and additional planes.

Several specifics tied to that upgraded version, including pricing, endurance ratings in drive-writes-per-day, and the exact generation number of the flash inside it, are not stated in Kioxia’s own cited announcement. Readers should treat the 10-million-IOPS figure as the confirmed, on-the-record spec from Kioxia, and the 20-million-plus figure as a reported upgrade that outlets are covering as the current headline number. Both numbers point in the same direction: Kioxia’s GP Series is scaling fast, and fast enough that storage watchers are now talking about it in the same breath as DRAM. The story has already spread beyond the original report, with outlets like Worldnews.com syndicating the same headline figures within hours.

Inside the Specs: PCIe 6.0, NVMe 2.2 and XL-FLASH Generation 2

The GP1 Series rides on PCIe 6.0, the newest PCI Express generation, paired with NVMe 2.2, the latest revision of the protocol SSDs use to talk to a host system. PCIe 6.0 roughly doubles the per-lane bandwidth of PCIe 5.0, which matters a lot for a drive whose entire reason for existing is moving huge volumes of small, random read requests without choking the bus.

Why the 512-Byte Block Size Matters

Most consumer SSD benchmarks you’ve seen quote 4KB random-read IOPS numbers, because that’s the block size most desktop operating systems and applications actually use. Kioxia instead measures the GP1 at 512 bytes, a much smaller block size that is harder to service quickly because the drive’s controller has to process far more individual operations to move the same amount of data. AI inference workloads, especially ones doing embedding lookups or sparse retrieval from huge parameter stores, tend to generate exactly this kind of small, random access pattern, which is why Kioxia is benchmarking at 512 bytes instead of the usual 4KB.

A simple way to see why this is a big deal: IOPS multiplied by block size gives you a rough throughput figure.

10,000,000 IOPS x 512 bytes  ≈ 4.88 GB/s of random-read throughput
20,000,000 IOPS x 512 bytes  ≈ 9.77 GB/s of random-read throughput

(Illustrative calculation based on Kioxia's confirmed 10M-IOPS spec
and the 20M-plus figure reported by TechRadar. Not an official
Kioxia bandwidth rating.)

That jump, from roughly 4.88 GB/s to nearly 9.77 GB/s of effective random small-block throughput, is the kind of leap that gets GPU cluster architects paying attention, because it starts closing the gap with memory tiers that used to be DRAM’s exclusive territory.

GPU Direct Access: Built for AI, Not Just Faster Storage

Kioxia describes the GP Series as the industry’s first super-high-IOPS PCIe SSD optimized for GPU direct access. In practice, that means the drive is designed to let a GPU pull data straight from flash without routing every request through the host CPU and system memory first. Cutting that detour out of the path reduces latency and frees up CPU cycles that would otherwise be spent shuttling data back and forth.

For large language model training and inference, this kind of direct path matters because GPUs are expensive, idle GPU time is wasted money, and anything that keeps a GPU fed with data instead of waiting on it pays for itself quickly at data-center scale. It’s the same basic logic that has driven interest in technologies like faster AI accelerator supply chains more broadly: the bottleneck in modern AI infrastructure increasingly isn’t raw compute, it’s getting data to the compute fast enough.

The Real Target: Flash as a Memory-Extension Tier for AI Servers

The phrase Kioxia uses in its own materials is telling: the GP1 is meant to provide “a flash-based memory-extension tier” that can extend High Bandwidth Memory. That’s a more modest and more precise claim than “SSD replaces DRAM,” and it reflects how storage vendors are actually trying to insert themselves into the AI memory hierarchy rather than overturn it.

Here’s the practical problem Kioxia is responding to. HBM and DRAM are fast but expensive and capacity-constrained, and both have been in short supply as AI buildouts accelerate. That shortage has rippled into adjacent markets too, including the kind of GDDR memory used in graphics cards affected by recent price increases. If a flash tier sitting just below HBM can absorb some of the memory capacity burden in an AI server, cloud operators and GPU cluster builders get a cheaper way to keep scaling model sizes without buying proportionally more DRAM. That’s the opportunity Kioxia is chasing, and it’s also why XL-FLASH generation 2 is being talked about in memory terms rather than purely storage terms.

How Kioxia’s Roadmap Scales Toward 100 Million IOPS

Kioxia’s own architecture claim, as cited in its announcement, lays out a scaling path from 10 million random-read IOPS in current GP Series products to a longer-term target of 100 million IOPS in future generations. That is a 10x jump from the company’s confirmed current spec, and it signals that the GP1 launched in August 2026 is explicitly positioned as a first step rather than a finished product line.

If the reported 20-million-plus figure for the upgraded drive is accurate, Kioxia would already be a fifth of the way to that 100-million-IOPS target less than three months after the GP1’s original debut. That pace of iteration looks less like a routine product refresh and more like a company racing to stake out ground in AI memory before rivals catch up.

Competition: Where Samsung, SK hynix, Western Digital and Micron Fit

Kioxia isn’t operating in a vacuum. Samsung, SK hynix, Western Digital and Micron all run their own high-performance enterprise SSD and advanced flash programs, and all four have talked publicly about pushing flash and flash-adjacent memory further into AI infrastructure roles traditionally held by DRAM and HBM. None of those companies has published a directly comparable 512-byte random-read IOPS figure for a shipping product in the window covered by this story, so a specs-for-specs comparison table isn’t possible without guessing. What can be said with confidence is that Kioxia’s move to frame its SSD roadmap explicitly around challenging DRAM, rather than simply outperforming other SSDs, puts pressure on every one of those rivals to articulate their own answer to the same AI-memory-shortage problem.

That competitive dynamic echoes what’s already playing out in AI chip supply chains, where market share swings have become a running story this year, including the shifts covered in Nvidia’s changing position in the China AI chip market. Memory and storage are following the same pattern: a scramble among a handful of large manufacturers to claim the piece of the AI stack that commands the best margins.

Historical Context: From Spinning Disks to Memory-Class Flash

It’s worth remembering how far storage performance has come in a relatively short span. A decade ago, enterprise storage was still dominated by spinning hard drives measured in a few hundred IOPS. The shift to NAND flash SSDs over PCIe brought that into the hundreds of thousands, then millions, of IOPS range. XL-FLASH itself is part of Kioxia’s longer effort, stretching back several product generations, to shrink latency closer to DRAM-like speeds while keeping flash’s cost and density advantages intact.

What’s different now is the demand side. AI training and inference workloads create sustained, massive, random small-block access patterns that older storage tiers were never designed to handle gracefully. That demand is the reason a company can go from “fast enterprise SSD maker” to “possible DRAM challenger” in the span of a few product cycles. The underlying flash technology evolved steadily; the market need for it just caught up all at once.

Market Impact: AI Infrastructure Spending Is the Backdrop

None of this is happening in isolation from the broader AI infrastructure buildout. According to IDC, organizations increased spending on compute and storage hardware infrastructure for AI deployments by 166% year over year in the second quarter of 2025, reaching $82.0 billion. Within that spending, AI infrastructure storage spending alone grew 20.5% year over year in the same quarter, with 48% of that spending coming from cloud deployments rather than on-premises buildouts.

IDC also tracked AI-centric storage reaching $1.76 billion in the third quarter of 2025, and the firm forecasts that global AI infrastructure spending will climb to $758 billion by 2029, with overall AI IT spending growing at a 31.92% compound annual rate, from $315.9 billion in 2024 to $1.262 trillion by 2029. That’s the financial backdrop against which a storage vendor’s IOPS claims actually mean something commercially. Every percentage point of that spending that shifts toward flash-based memory-extension tiers instead of pure DRAM is real revenue for whoever gets there first with a credible product.

Metric	Figure	Period
AI compute/storage hardware spending	$82.0 billion (+166% YoY)	Q2 2025
AI infrastructure storage spending growth	+20.5% YoY	Q2 2025
Share of AI storage spend from cloud deployments	48%	Q2 2025
AI-centric storage market size	$1.76 billion	Q3 2025
Global AI infrastructure spending forecast	$758 billion	By 2029
AI IT spending compound annual growth rate	31.92% ($315.9B to $1.262T)	2024-2029

Figures cited from IDC research on AI infrastructure and storage spending.

KIOXIA GP1 Series Specs at a Glance

To keep the confirmed facts separate from the reported upgrade, here’s what Kioxia’s own GP1 announcement actually states, alongside the company’s stated long-term roadmap target.

Attribute	KIOXIA GP1 Series Detail
Manufacturer	Kioxia Corporation
Series	KIOXIA GP1 Series (first product in the GP Series)
Interface	PCIe 6.0
Protocol	NVMe 2.2
Flash memory	KIOXIA XL-FLASH generation 2
Confirmed random-read performance	Up to 10 million IOPS at 512-byte block size
Long-term roadmap target	Up to 100 million IOPS in future generations
Memory role	Flash-based memory-extension tier to supplement HBM
Primary use case	AI applications and GPU direct access
Announcement date	August 4, 2026
Pricing	Not disclosed
Endurance rating	Not disclosed

Specifications drawn from Kioxia’s official GP1 Series announcement, as cited by the company.

What This Means for Enterprise and Cloud Buyers

For cloud providers and enterprise buyers building out AI clusters, the practical question isn’t whether 10 million or 20 million IOPS sounds impressive. It’s whether a flash-based memory-extension tier can be slotted into existing GPU server designs without a redesign, and whether the economics favor buying more of this kind of SSD over buying more DRAM or HBM capacity. Given how tight memory supply has been across the industry, including the pricing pressure already visible in GPU memory markets, a credible flash alternative that eases even a slice of that demand is worth serious evaluation for anyone planning large GPU deployments into 2027.

That said, buyers should wait for Kioxia or a verified third-party benchmark to confirm the 20-million-IOPS figure, along with pricing and endurance numbers, before treating the upgraded drive as a finalized product. Procurement teams have been burned before by headline specs that didn’t survive contact with production hardware, and this is a case where the gap between “confirmed by the manufacturer” and “reported by the press” is still wide enough to matter.

Predictions: Where High-IOPS Flash Goes From Here

Kioxia will likely publish an official follow-up announcement confirming (or correcting) the 20-million-IOPS figure now circulating in press coverage, since the gap between its own 10-million spec and the reported number is too large to go unaddressed.
Expect Samsung, SK hynix and Western Digital to respond with their own GPU-direct, high-IOPS SSD announcements within the next two to three quarters, following the same AI-memory-extension framing.
Pricing for GP Series drives will likely land at a premium well above standard enterprise NVMe SSDs, given the specialized XL-FLASH generation 2 memory and the narrow AI-infrastructure target market.
If Kioxia’s 100-million-IOPS roadmap target holds, later GP Series generations are unlikely to arrive before 2028, given the scale of the engineering jump from the current 10-million baseline.
AI infrastructure spending growth, as tracked by IDC, will keep pulling storage and memory vendors toward GPU-adjacent product categories, making “memory-extension” positioning a more common pitch across the SSD industry through 2027.

Frequently Asked Questions

What is the KIOXIA GP1 Series?

The KIOXIA GP1 Series is a line of PCIe 6.0 NVMe SSDs that Kioxia Corporation announced on August 4, 2026, as the first product in its new GP Series of super-high-IOPS drives built for AI applications and GPU direct access.

How many IOPS does the Kioxia GP1 SSD deliver?

Kioxia’s own official announcement states up to 10 million random-read IOPS at a 512-byte block size. A later report from TechRadar describes an upgraded edition of the drive exceeding 20 million IOPS under the same test conditions, though that higher figure has not been confirmed in Kioxia’s original cited press material.

What’s the difference between the 10 million and 20 million IOPS figures?

The 10-million figure is the confirmed, on-the-record spec from Kioxia’s official GP1 announcement. The 20-million-plus figure is a reported number for a faster edition of the drive, covered by TechRadar, and includes details, like a third-generation XL-FLASH design, that are not part of Kioxia’s original cited release.

What is KIOXIA XL-FLASH generation 2?

XL-FLASH generation 2 is the low-latency flash memory that powers the GP1 Series. It’s designed to deliver faster random-read performance than standard NAND flash, which is part of how Kioxia reaches its million-plus IOPS figures at small block sizes.

Can SSDs really replace DRAM in AI servers?

Not entirely, at least not yet. Kioxia positions the GP1 as a flash-based memory-extension tier that supplements High Bandwidth Memory rather than replacing DRAM outright. The goal is to absorb some of the memory capacity burden in AI servers, not to eliminate DRAM from the architecture.

When did Kioxia announce the GP1 Series?

Kioxia’s official GP1 Series announcement is dated August 4, 2026. Reports of an upgraded, faster edition of the drive surfaced in October 2026.

What’s next on Kioxia’s GP Series roadmap?

Kioxia has stated an architecture goal of scaling from 10 million random-read IOPS in current GP Series products to a longer-term target of 100 million IOPS in future generations, though no firm timeline for that target has been disclosed.

Is the KIOXIA GP1 Series available to buy now?

Kioxia’s cited announcement does not disclose pricing or a retail availability date, and these GP Series drives are aimed at AI infrastructure and enterprise GPU deployments rather than consumer purchase.

Kioxia’s SSD Hits 20M IOPS, Takes Aim at DRAM [2026]