Kioxia GP1 Hits 10 Million IOPS, but the Hardware Still Has More to Prove

Kioxia demonstrated its GP1 Series at 10 million input/output operations per second, giving Google News readers a concrete sign that AI storage is changing. The figure is impressive, yet it represents an early device demonstration rather than a shipping product validated inside production AI clusters.

The GP1 belongs to Kioxia’s broader GP Series, a family of SSDs designed for direct, fine-grained access from GPUs. Kioxia plans to offer evaluation samples to selected customers by the end of 2026. Its longer-term target exceeds 100 million IOPS, leaving a sizable gap between the latest demonstration and the intended architecture.

That gap is the real story. Kioxia is not merely trying to build a faster enterprise drive. It wants flash to become an extension of expensive GPU memory, competing with DRAM-centered designs for selected AI data. SK hynix, Samsung, Solidigm, and other storage suppliers are pursuing related paths, while Nvidia is defining the systems that could make them useful.

Google News Headlines Miss the Two Different Kioxia Milestones

Kioxia has demonstrated a 10 million IOPS device, while its 100 million IOPS figure remains a next-generation system target.

The distinction matters because several Kioxia disclosures now circulate together. Their figures describe different stages of the company’s roadmap, not one finished product with a single verified specification.

Kioxia first outlined a super-high-IOPS SSD during its 2025 corporate strategy presentation. The company described a device using XL-FLASH and a new controller, with more than 10 million IOPS expected during 2026. It also positioned direct GPU access as a way to reduce CPU involvement in the data path.

In March 2026, Kioxia formally announced development of the GP Series. Its GP Series release described GPU-accessible flash, 512-byte requests, and lower energy use per operation than conventional Kioxia TLC drives. The announcement did not publish a complete commercial specification sheet.

Kioxia then used its June investor event to expand the ambition. The company said future GP products would address AI systems requiring more than 100 million IOPS. Its Investor Day update presented that performance alongside XL-FLASH, a dedicated controller, and support for Nvidia Storage-Next.

Those claims do not mean that the GP1 demonstration already reached 100 million IOPS. The more defensible reading is that GP1 establishes the first step at about 10 million. Later GP generations and larger system configurations are expected to pursue the higher target.

Kioxia also uses “GP Series” in official English materials, while some FMS 2026 coverage identifies the demonstrated first-generation device as GP1. Readers should not treat GP1, GP Series, and the 100 million IOPS roadmap as interchangeable labels.

This is one reason aggregation can flatten important technical context. A Google News headline captures the announcement, but not the boundary between an emulator, demonstrated hardware, evaluation samples, and qualified production drives.

The product timeline remains equally important. Kioxia says selected customers should receive GP Series evaluation samples by the end of 2026. Evaluation samples let system builders test integration, firmware, workloads, and reliability. They are not evidence of broad availability or deployment at scale.

That means buyers still lack several ordinary purchasing details. Kioxia has not provided a final capacity range, endurance specification, sustained workload profile, form factor matrix, or general-availability date for GP1. It has also not published broad third-party benchmark results.

The news is real, but its scope is narrower than the most dramatic number suggests. Kioxia has shown that its specialized flash architecture can enter a performance class beyond conventional enterprise SSDs. It has not yet shown that the technology can replace a meaningful share of HBM or DRAM in production inference.

Why AI Systems Suddenly Need So Many Small Operations

The GP1 is optimized for frequent, tiny requests because agentic inference stresses storage differently from training or conventional file access.

IOPS measures how many input/output operations a device completes each second. It does not measure the amount of data transferred by itself. A drive can post high IOPS with small requests while delivering less total bandwidth than another device moving large sequential blocks.

Traditional enterprise SSD marketing often emphasizes sequential throughput and four-kilobyte random operations. Those metrics fit databases, virtual machines, analytics, and general server storage. GPU-directed AI access can involve smaller pieces of data requested at much higher concurrency.

Kioxia’s design uses 512-byte access granularity, meaning each operation can retrieve a smaller unit than a conventional four-kilobyte request. That difference can reduce unnecessary data movement when the useful item occupies only a fraction of a larger block.

The company pairs that access pattern with XL-FLASH, its low-latency storage-class memory. XL-FLASH uses single-level cell storage, which keeps one bit in each memory cell. That approach favors latency and endurance over the greater density available from TLC or QLC flash.

Kioxia’s own technical explanation says the first GP generation is intended to support 100 million 512-byte random-read operations. It also places second-generation sampling in 2027. Those targets remain company projections until customers test representative hardware.

One target workload is retrieval-augmented generation, or RAG. This method retrieves external information before a model constructs its answer. Large vector indexes can involve many small, irregular reads, particularly when databases keep only their hottest data in memory.

Another target involves model parameters that cannot all remain in GPU high-bandwidth memory. A mixture-of-experts model, for example, activates selected portions of a larger network for each token. Fast storage could hold less frequently used portions if software predicts and retrieves them before the GPU stalls.

KV cache presents a related but more complicated opportunity. A key-value cache stores intermediate attention data created while a language model processes a conversation. Reusing that data avoids repeating earlier computation, but long prompts and concurrent sessions can consume considerable memory.

GPUs need immediate access to the hottest cache entries. Less active context can move through a hierarchy containing HBM, host memory, local SSDs, and shared storage. Nvidia’s KV cache manager reflects this tiered approach by coordinating cache blocks across several memory and storage locations.

Not every cache workload favors tiny reads, however. Software can combine many tokens into larger blocks before moving them to storage. Samsung described 33-megabyte cache files in one inference storage study, making sequential bandwidth important in that tested configuration.

This variation prevents one IOPS number from predicting application performance. Vector search, parameter retrieval, model loading, and KV cache offload can produce different block sizes, queue depths, read-to-write ratios, and latency requirements.

The application also determines whether the CPU sits in the data path. GPU-initiated access can reduce software overhead and unnecessary transfers, but it needs support across drivers, networking, security, orchestration, and storage firmware.

Kioxia is therefore selling an architectural idea as much as a device. Its thesis is that AI servers need a storage tier built for GPU access patterns rather than a faster version of familiar CPU-oriented NVMe storage.

The Real Contest Is Flash Expansion Versus Expensive Memory

Kioxia must show that flash can extend GPU-accessible capacity without making accelerators wait for every important request.

High-bandwidth memory sits close to a GPU and supplies far lower latency with far greater bandwidth than NAND flash. It is also limited in capacity and expensive to deploy. Those properties make HBM valuable for active model data and computation, but difficult to scale as a universal repository.

DRAM adds another tier with more capacity and lower performance than HBM. It still costs more per unit of capacity than NAND. Large inference systems already divide data among these layers, moving information according to urgency, reuse, and available space.

GP1 does not erase the physical difference between memory and storage. It tries to make flash useful for data that can tolerate microseconds rather than nanoseconds, provided the device handles enough simultaneous requests.

This creates the article’s primary conflict: specialized flash expansion versus continued reliance on HBM and DRAM. Kioxia wins only if its SSD allows systems to serve more useful work from each expensive GPU without adding damaging latency.

A single 10 million IOPS drive can look remarkable beside conventional enterprise SSDs. The relevant comparison changes at the system level. Multiple GPUs can generate enormous aggregate demand, while each stalled request can leave costly compute resources idle.

Kioxia’s custom controller is central to the argument. The controller must schedule many small operations, communicate efficiently with GPUs, and manage XL-FLASH without turning software overhead into the new bottleneck. Power use per operation also matters because AI racks already face strict electrical and cooling limits.

Nvidia Storage-Next supplies the broader architecture. It asks storage vendors to build devices for GPU-initiated workloads and treats flash as part of GPU-accessible memory. Kioxia’s GP Series is one supplier response, not an isolated proprietary experiment.

The idea differs from Nvidia’s Context Memory Storage approach, which concentrates on KV cache and high-bandwidth movement through a tiered memory system. Kioxia positions its CM Series for that bandwidth-oriented role. It reserves the GP Series for lower-latency, high-IOPS access and RAG workloads.

That portfolio separation is revealing. Kioxia does not claim that one exotic SSD can handle every AI storage pattern. It offers TLC drives for bandwidth and endurance, XL-FLASH for fine-grained access, and QLC products for high capacity.

The GP1’s lower density creates an economic tradeoff. SLC-based XL-FLASH dedicates more physical flash cells to a given capacity than TLC or QLC NAND. Buyers therefore need enough application value from lower latency to justify using it.

The return can come from higher GPU utilization, greater model capacity, more concurrent sessions, or faster retrieval. Yet none of those gains follows automatically from a drive-level IOPS result. Software must place the correct data on GP1 and request it early enough.

Developers should view the SSD as another tier with explicit scheduling requirements. A poorly designed pipeline could move data repeatedly, create cache churn, or overwhelm the interconnect. Those failures would reduce the value of the faster media.

The architecture becomes more persuasive when workloads have a large working set with identifiable hot and warm regions. HBM can hold the hottest information, while GP1 holds warm data that remains performance-sensitive. High-capacity TLC or QLC storage can retain colder material.

This hierarchy does not replace memory. It changes how much data must occupy the most expensive layer at every moment. That narrower proposition is technically credible, but buyers need application benchmarks to quantify it.

Kioxia Is Entering a Race, Not Creating a Category Alone

Competitors are pursuing the same storage opportunity with different media, interfaces, and performance priorities.

SK hynix has described an AIN P product aimed at fine-grained AI inference access. The proposed design uses 512-byte operations and targets about 50 million IOPS over PCIe 6.0, followed by a path toward 100 million. Its expected timing extends into 2027.

That roadmap closely resembles Kioxia’s central thesis. Both companies expect AI inference to generate enough small random operations to justify controllers and flash configurations designed for a new performance class.

The comparison also shows why Kioxia’s early demonstration matters. A functioning GP1 device at 10 million IOPS gives the company hardware that customers can begin evaluating before every roadmap reaches its final target.

It does not settle the contest. SK hynix could compete through integration with its broader memory portfolio. Samsung can combine enterprise SSDs, DRAM, HBM, and system validation. Solidigm has emphasized large-capacity storage and software-assisted KV cache or vector retrieval.

Conventional enterprise SSDs continue improving as well. Samsung lists up to 3.3 million random-read IOPS and 14.5 gigabytes per second for its PM1753. Those figures trail GP1’s headline IOPS, but the Samsung device also offers documented enterprise features and workload testing.

Kioxia’s existing CM7 comparison material lists 1.4 million IOPS for a PCIe 5.0 enterprise drive. Against that baseline, a 10 million IOPS GP1 represents a substantial device-level increase. The gap helps explain why specialized media and controllers are receiving attention.

Historical context introduces another competitor, even though it is no longer actively developed. Intel Optane used 3D XPoint media to provide latency and endurance between DRAM and NAND SSDs. Its commercial decline proved that technical advantages do not guarantee a sustainable market.

Optane faced difficult economics, limited adoption, and the challenge of fitting a new memory tier into existing systems. GP1 benefits from NAND manufacturing experience and Nvidia’s architectural support, but it still faces integration and demand risks.

Industry analysis also separates raw performance from useful deployments. ServeTheHome’s assessment argues that design wins depend heavily on Nvidia pushing Storage-Next into future systems. That dependency gives Nvidia significant influence over adoption.

The software layer can become another competitive boundary. Storage vendors need compatibility with inference runtimes, vector databases, orchestration systems, and GPU communication frameworks. Hardware that requires extensive custom work will appeal mainly to the largest operators.

Open software support can broaden the market. Clear interfaces would let infrastructure teams compare GP1 with ordinary NVMe arrays, remote memory, compressed caches, or recomputation. Closed integrations could produce strong results in selected systems while limiting wider adoption.

Kioxia also needs to explain where GP1 sits beside its CM9 family. CM9 offers a documented capacity of 25.6 terabytes and endurance of three drive writes per day. GP1 currently emphasizes access granularity and IOPS without an equally complete public specification.

That difference suggests CM9 is the nearer-term option for bandwidth-heavy KV cache deployments. GP1 is the more experimental option for tiny requests and GPU memory extension. Combining their claims would make both products harder to evaluate.

Google News coverage can make the market look like a simple SSD speed contest. The actual competition involves memory placement, interconnects, firmware, software scheduling, endurance, capacity, power, and total system economics.

What the 10 Million IOPS Figure Does Not Prove

A device demonstration cannot establish production latency, endurance, power efficiency, or gains in tokens per second.

Benchmark numbers depend on workload settings. IOPS changes with request size, queue depth, read percentage, data pattern, overprovisioning, and the number of workers submitting commands. A peak result may not describe sustained behavior under mixed AI traffic.

Kioxia says GP1 supports 512-byte access, which naturally produces more operations for a given data volume than four-kilobyte requests. Buyers should compare both IOPS and transferred bytes before drawing conclusions about total throughput.

Ten million operations at 512 bytes represent about 5.12 gigabytes of requested data each second before protocol overhead. The calculation illustrates why an exceptional IOPS number can coexist with moderate aggregate bandwidth. The device is optimized for granularity, not merely bulk transfer.

Latency distribution is equally important. An average can hide occasional slow requests, while tail latency measures the worst experiences among a high percentage of operations. AI services running many parallel requests often care about those outliers.

A delayed cache fetch can stall a generation step even if most reads finish quickly. Buyers therefore need percentile latency at realistic queue depths, not only average latency under a synthetic random-read test.

Endurance remains another open question. KV cache systems write new state, delete expired sessions, and move data across tiers. Vector indexes also change as organizations update documents or embeddings. XL-FLASH should offer advantages over denser NAND, but GP1 needs a published workload rating.

Capacity will determine the device’s role. A small, fast SSD can serve as a warm cache, while a larger device can hold more model parameters or retrieval data. Without final capacities, architects cannot calculate drive counts, rack space, or failure exposure.

Power consumption per operation is one of Kioxia’s stated advantages over its conventional TLC SSDs. The company has not yet supplied enough public data to compare full-device power under representative workloads. Cooling requirements and idle behavior also affect deployment.

Reliability features need scrutiny because direct GPU access changes the data path. Operators still require error handling, access controls, isolation, telemetry, firmware management, power-loss protection, and predictable recovery after a device failure.

Security becomes particularly important in shared infrastructure. Cached prompts or retrieved enterprise data can contain sensitive information. GPU-directed storage must preserve tenant isolation and prevent stale data from crossing workload boundaries.

Software overhead can erase some media-level gains. Drivers, memory registration, address translation, request batching, and cache management all consume resources. A fast SSD cannot help when a runtime fails to issue requests efficiently.

System benchmarks should measure time to first token, tokens per second, concurrent users, GPU utilization, and energy per completed request. Those metrics connect storage behavior to the service an AI operator actually sells.

A credible test should also compare GP1 against more DRAM, conventional NVMe drives, recomputation, and remote cache systems. Each alternative carries different capital, power, capacity, and operational costs.

The strongest evidence would come from customer workloads that exceed HBM capacity and repeatedly access fine-grained data. If GP1 increases useful GPU work while holding response latency steady, Kioxia’s architectural claim gains support.

Until then, the 10 million IOPS demonstration proves a narrower point. Kioxia has built a specialized flash device capable of unusually high small-block operation rates under demonstrated conditions. It does not prove that every AI application needs the product.

Three Signals Will Decide Whether GP1 Becomes AI Infrastructure

Sample specifications, application benchmarks, and platform adoption will determine whether GP1 moves beyond an impressive FMS demonstration.

The first signal is Kioxia’s evaluation-sample package. The company expects selected customers to receive GP Series samples by the end of 2026. Those devices should reveal final capacities, interfaces, form factors, endurance, power limits, and sustained performance.

A delay would weaken the current timeline. On-time sampling with complete documentation would strengthen the case that GP1 is becoming a deployable product rather than a technology showcase.

The second signal is an end-to-end benchmark from Kioxia, Nvidia, or a customer. The most useful result would compare the same model and server with different memory hierarchies. It should report GPU utilization, response latency, throughput, and energy.

A drive-level IOPS chart will not be enough. The central promise is better AI system efficiency, so validation must show more completed inference work or larger usable models under controlled conditions.

Kioxia should also disclose whether its demonstrated result uses one device, multiple controllers, an emulator, or a complete Storage-Next node. That context will help buyers estimate the hardware required for higher system-level targets.

The third signal is adoption of Nvidia Storage-Next by server manufacturers, cloud providers, and inference software projects. GP1 depends on more than a compatible PCIe slot. It needs a stable path for GPUs to address, schedule, secure, and recover data stored on flash.

Support from a major server vendor would strengthen Kioxia’s position. Integration into common inference runtimes would matter even more, because developers generally adopt new hardware through software they already use.

Competitor timing will influence these signals. If SK hynix samples its high-IOPS products quickly, buyers will gain leverage and comparison data. If conventional TLC drives close much of the performance gap, specialized SLC economics become harder to defend.

The 100 million IOPS target deserves separate treatment. Kioxia describes it as a requirement for next-generation AI systems and a GP Series objective. The company has not established that a single shipping GP1 device sustains that rate today.

Future Google News updates should therefore be read against three questions. Is the number measured or projected? Does it describe one drive or a system? Does the benchmark improve an actual AI workload?

For developers, this development matters because memory placement is becoming part of inference engineering. Teams may need to profile which model data, cache blocks, and retrieval indexes deserve HBM, DRAM, specialized flash, or ordinary storage.

Enterprise buyers should ask for workload traces and application results before reserving capacity. They should also examine failure behavior, observability, security controls, and software portability. Peak IOPS alone cannot answer those operational questions.

Knowledge workers will not purchase GP1 drives directly, but they can still feel the outcome. Better memory utilization can support longer sessions, more concurrent agents, and broader retrieval without requiring proportional growth in GPU memory.

Kioxia has earned attention by showing a first-generation device around 10 million IOPS and articulating a path toward a much larger target. The next step is less theatrical and more important: proving that those operations keep GPUs productively occupied.

Watch the evaluation samples, not only the headline. Look for measured application gains, clear system configurations, and commitments from platform partners. Those signals will show whether the GP1 becomes a practical AI memory tier or remains an exceptional SSD benchmark.