Kioxia Expands Its AI Storage Push With UFS 5.0 and Next-Generation NAND

Kioxia has hit Google News with two ambitious storage moves, yet its UFS 5.0 numbers do not surpass Samsung’s announced mobile performance. The Japanese memory manufacturer is sampling new embedded storage while preparing faster NAND for future PCIe 6.0 data center products. Together, these developments show how Kioxia intends to compete across both ends of the AI market.

The important conflict is not one specification against another. Kioxia is challenging Samsung and SK hynix through a broader NAND strategy. It wants its flash inside smartphones running local language models and servers feeding data to AI accelerators. That approach differs from the industry’s intense focus on high-bandwidth memory, or HBM, which places fast DRAM beside processors.

However, the original headline compresses several separate developments into a single claim. UFS 5.0 targets mobile and edge devices. Kioxia’s tenth-generation BiCS flash targets performance-sensitive storage, including future PCIe 6.0 enterprise drives. Its experimental high-bandwidth flash work occupies another category entirely.

Kioxia therefore is not introducing one product that simultaneously combines UFS 5.0, PCIe 6.0, and every form of “AI NAND.” It is assembling a portfolio around a common bet. NAND must move closer to computation because AI models and retrieval databases are growing too large for available DRAM.

That bet matters, but announcements and commercial victories are not equivalent. Samsung has published slightly faster UFS 5.0 specifications and a firm mass-production window. SK hynix has its own AI-NAND family and a high-bandwidth flash partnership with SanDisk. Kioxia must turn samples, demonstrations, and component technology into customer-qualified products.

What Kioxia Actually Announced

Kioxia’s immediate product is a pair of UFS 5.0 samples, not a finished PCIe 6.0 AI drive.

On July 29, Kioxia America announced commercial samples of UFS 5.0 embedded flash in 512GB and 1TB capacities. Universal Flash Storage, or UFS, is the compact storage standard commonly used in premium smartphones and other mobile devices.

The company says its devices can reach sequential reads of 10GB/s and sequential writes of 9GB/s. The 9GB/s write figure applies to the 1TB model. Kioxia expects mass production to begin by the end of 2026.

Its UFS 5.0 samples use the MIPI M-PHY 6.0 physical layer and UniPro 3.0 protocol. These interface technologies govern how data travels between embedded storage and a device’s processor.

The underlying interface supports a theoretical 46.6Gb/s per lane. A dual-lane implementation provides about 10.8GB/s of effective theoretical performance. Actual product speeds remain below that ceiling because controllers, flash media, thermal limits, and software introduce overhead.

Kioxia pairs an internally designed controller with its eighth-generation BiCS 3D flash. BiCS stacks memory cells vertically, allowing manufacturers to increase capacity without relying only on smaller horizontal features. The package measures 7.5 by 13 millimeters and includes improved thermal management, according to Kioxia.

The announced use cases extend beyond flagship phones. Kioxia also lists tablets, gaming systems, extended-reality hardware, robots, security cameras, and industrial edge equipment. Each category can generate or process large datasets without sending every operation to a cloud server.

This local processing angle explains the AI language surrounding the launch. A phone must first move model parameters from persistent flash into faster working memory before its processor can perform inference. Slow storage can lengthen application launches and delay the first generated response.

Kioxia says UFS read performance has become a limiting factor as local models grow. That statement is technically plausible, but storage represents only one part of the system. Model architecture, quantization, DRAM capacity, memory bandwidth, processor throughput, and software optimization also affect response time.

The PCIe 6.0 connection comes from a different Kioxia program. The company and SanDisk have developed tenth-generation BiCS flash with a 4.8Gb/s NAND interface. That component is intended to support future data center storage, including drives designed around faster host interfaces.

PCI Express, or PCIe, links storage and accelerators to a server’s processor. PCIe 6.0 doubles the raw transfer rate of PCIe 5.0 to 64 gigatransfers per second per lane. A compliant NAND component alone does not create a complete PCIe 6.0 SSD.

That distinction is central to interpreting the Google News headline. Kioxia has credible building blocks across mobile and enterprise storage. It has not announced one unified device that erases the boundaries between those markets.

Why Google News Is Connecting UFS 5.0 With AI

Storage is becoming part of AI performance because larger local models must repeatedly cross the boundary between flash and working memory.

Kioxia has described a typical on-device workflow in unusually concrete terms. A language model remains in UFS while inactive. The device loads its parameters into DRAM when an application needs inference, then the system-on-chip performs the calculations.

Kioxia estimates that local models containing three to four billion parameters can occupy roughly 3GB to 4GB after eight-bit quantization. Quantization reduces the numerical precision of model weights, lowering memory requirements at some cost to accuracy.

UFS 4.0 and 4.1 can load models of that size in less than one second under favorable conditions, Kioxia says. Larger models increase the transfer time and delay the first token. A faster storage interface becomes more valuable as phones hold more capable models.

The company’s on-device AI analysis argues that UFS 5.0 can support model files around 10GB. This is a company estimate, not an independent guarantee of application performance. A model’s size does not reveal its quality, processor demands, or usable context length.

Kioxia also wants smartphones to hold retrieval-augmented generation databases. Retrieval-augmented generation, or RAG, searches an external collection before a model answers. The retrieved material supplies information that the model did not retain in its trained parameters.

A local RAG collection could contain documents, messages, images, or application data. Keeping that index on the device can reduce cloud dependence and support private workflows. It also creates irregular reads that differ from copying one large file sequentially.

That difference exposes a limitation in headline specifications. Sequential throughput measures long, orderly transfers under controlled conditions. AI retrieval can involve many smaller requests, metadata lookups, decompression steps, and contention with other applications.

Random-read latency and sustained performance may matter more than a peak sequential figure in those situations. Thermal throttling can also reduce speed after a device heats up. Phone makers will need to test complete systems rather than compare storage labels alone.

Kioxia has developed AiSAQ, software that searches vector data while the database remains on an SSD instead of occupying DRAM. The company says it has technically verified a possible smartphone implementation. It has not announced a commercial phone using that combination.

The appeal is easy to understand. Mobile DRAM is expensive, consumes power, and competes for limited board space. Flash offers much greater capacity, although it remains slower and behaves differently under sustained workloads.

Moving selected AI data from DRAM to flash therefore involves a tradeoff. Manufacturers gain capacity and persistence but accept higher latency. Faster UFS narrows the penalty without turning NAND into DRAM.

This is why the story has traveled through Google News and technology feeds. The change is not simply that phone storage can copy files faster. Storage vendors are positioning flash as an active layer in the AI memory hierarchy.

That hierarchy stretches from processor caches and HBM to conventional DRAM, specialized flash, and bulk SSD capacity. Each layer balances speed, capacity, energy use, endurance, and cost. Kioxia wants NAND to handle more work without claiming it can replace every faster tier.

Samsung Still Leads the Published UFS 5.0 Numbers

Kioxia has joined the UFS 5.0 race, but Samsung currently publishes the stronger performance claim and the earlier production schedule.

Samsung announced its UFS 5.0 solution on June 23, more than a month before Kioxia’s commercial-sample announcement. Samsung claims sequential reads of 10.8GB/s and writes of 9.5GB/s. Both figures exceed Kioxia’s stated 10GB/s and 9GB/s maximums.

The comparison does not establish how either product performs inside a retail phone. Both companies are publishing controlled specifications rather than independent device benchmarks. Controller firmware, capacity, workload duration, cooling, and host integration can reshape the result.

Still, Samsung’s published lead weakens any simple claim that Kioxia has overtaken its Korean rival. Samsung also says its product improves power efficiency by more than 40 percent compared with the company’s UFS 4.1 solution.

Its package measures 7.5 by 13 by 0.9 millimeters, which Samsung describes as 16.7 percent smaller than its predecessor. Planned capacities reach 1TB. The company expects mass production during the fourth quarter of 2026.

Kioxia expects production by year-end, leaving the companies with closely overlapping schedules. Winning the specification announcement matters less than securing qualification inside flagship devices. Mobile manufacturers typically validate storage well before shipping products at scale.

Samsung has an additional structural advantage. It sells memory, processors, displays, image sensors, and finished Galaxy devices. That range gives it opportunities to coordinate component development and test new storage against real mobile workloads.

Kioxia remains a focused flash supplier. Its specialization can support faster design decisions and deep controller expertise. However, it depends more heavily on external device makers to convert component performance into a visible consumer experience.

The strongest version of Kioxia’s case is therefore not “faster than Samsung.” It is that customers now have another advanced UFS 5.0 supplier with an internal controller and internally developed NAND. A credible second source can matter to phone makers managing supply, qualification, and negotiating leverage.

UFS 5.0 also arrives before software developers have fully exploited its bandwidth. Faster model loading is useful, but an application must know when to prefetch weights, which data to cache, and how to limit memory duplication.

Operating-system support will influence those decisions. So will mobile chipsets capable of feeding storage data into neural-processing engines without creating new bottlenecks. A fast UFS package cannot compensate for a narrow pathway elsewhere.

Power deserves similar scrutiny. Samsung published a specific generational efficiency claim, while Kioxia said its efficiency improved without providing a percentage in the U.S. announcement. Neither statement replaces measurements from production phones.

Higher peak speed can reduce energy by finishing a transfer sooner. It can also raise instantaneous power and heat. The balance depends on workload design, controller behavior, and the time a device spends at maximum throughput.

For buyers, the first meaningful comparison will come from matched devices or engineering platforms. Those tests should include model-loading time, first-token latency, sustained retrieval, surface temperature, and battery consumption.

Until those results arrive, Samsung holds the cleaner marketing position. Kioxia has competitive specifications and comparable capacity, but its published numbers do not support an outright performance victory.

PCIe 6.0 Is a Roadmap, Not a Shipping Kioxia SSD

Kioxia’s enterprise argument rests on faster NAND that can feed future PCIe 6.0 systems, not on a broadly available PCIe 6.0 product.

Kioxia and SanDisk first detailed their tenth-generation 3D flash technology at the 2025 International Solid-State Circuits Conference. The design uses 332 memory layers and a 4.8Gb/s NAND interface.

The companies said that interface is 33 percent faster than their eighth-generation flash. They also reported 10 percent lower input power and 34 percent lower output power. Bit density improves by 59 percent through additional layers and a redesigned floor plan.

These are component-level claims. They describe communication between NAND dies and a controller, plus the density of information stored on silicon. They do not describe the final speed, endurance, latency, or capacity of a complete SSD.

The BiCS technology uses CMOS directly bonded to array, or CBA. Kioxia and SanDisk manufacture the control circuitry and memory-cell array on separate wafers, then bond them together.

Separate production lets engineers optimize each wafer for a different job. Logic can use a process suited to fast switching, while the memory array prioritizes density and cell characteristics. It also creates manufacturing and yield challenges at the bonding stage.

The new design separates command and address traffic from data transfers. That arrangement lets the chip receive instructions while moving data, reducing idle periods. Power-isolation techniques lower energy consumption during input and output operations.

Faster NAND is necessary because PCIe 6.0 expands the pipe between an SSD and its host. A four-lane connection offers roughly twice the theoretical bandwidth of PCIe 5.0. Controllers need many fast NAND channels to keep that host interface busy.

Kioxia’s own PCIe specifications list PCIe 6.0 at 64 gigatransfers per second per lane. The standard also changes signaling and error-management methods to maintain reliability at that speed.

A drive cannot reach useful PCIe 6.0 performance by replacing only its flash. It requires a compatible controller, firmware, connector, server platform, power design, and thermal system. Host processors and operating systems must also support the interface.

This is why “PCIe 6.0 AI NAND” works better as a roadmap description than as a product category. Kioxia is preparing media capable of serving future drives. Customer qualification, controller readiness, and server adoption will decide when the promise becomes commercially relevant.

The AI connection comes from data movement. Training and inference clusters continually fetch model checkpoints, embeddings, datasets, and cached results. Faster SSDs can shorten periods when expensive accelerators wait for input.

Yet many AI bottlenecks remain outside storage. Network congestion, preprocessing, software orchestration, accelerator memory capacity, and database design can dominate application latency. Doubling one interface does not double an entire system’s performance.

Kioxia must therefore demonstrate more than a benchmark peak. It needs sustained throughput under mixed reads and writes, predictable tail latency, acceptable endurance, and manageable power. Enterprise buyers also expect error handling and fleet-level reliability data.

Those requirements favor established suppliers, including Kioxia, Samsung, Solidigm, and Micron. They also slow the transition between interface generations. Data center customers rarely replace qualified storage solely because a faster link becomes available.

Kioxia’s advantage is timing. Its NAND roadmap anticipates the bandwidth needs of PCIe 6.0 before those drives become commonplace. The risk is that controllers, servers, or workloads reach broad adoption later than its component schedule.

SK Hynix Is Building a Different AI-NAND Challenge

SK hynix is not waiting for conventional SSD interfaces to define AI storage, which makes it a more complicated opponent than the headline suggests.

SK hynix describes an “AI-NAND Family” with products specialized for different workloads. AIN-P targets faster enterprise SSDs. AIN-B uses through-silicon vias, or TSVs, to create wider connections through vertically stacked dies.

The company is commercializing AIN-B as high-bandwidth flash, or HBF, through a partnership with SanDisk. HBF attempts to give accelerators much more flash capacity through an architecture inspired by stacked high-bandwidth memory.

SK hynix says the program includes logic processing in a base die. Logic below stacked flash can coordinate access and reduce some data movement. However, NAND latency and endurance still differ fundamentally from DRAM-based HBM.

The company’s AI-NAND strategy therefore attacks a different level of the memory hierarchy. UFS 5.0 improves local storage for mobile devices. PCIe 6.0 accelerates storage attached through a conventional server interface. HBF aims to sit closer to accelerators.

Kioxia has also explored high-bandwidth flash concepts. Its research strengthens the broader thesis that AI systems need an intermediate capacity tier between HBM and ordinary SSDs. It does not mean the industry has agreed on one architecture or standard.

This competition creates several possible paths. Conventional enterprise SSDs can become faster and more efficient. Compute Express Link devices can add capacity through a shared fabric. HBF can attach large NAND pools nearer to accelerators. Software can also stream model data more intelligently from existing storage.

These routes can coexist because AI workloads differ. Training a frontier model places different demands on storage than serving a recommendation system. A smartphone assistant faces tighter power and thermal limits than a rack-scale inference cluster.

Kioxia’s portfolio approach reduces dependence on one route. It has mobile UFS, enterprise NAND, SSD controllers, retrieval software, and experimental high-bandwidth technologies. That breadth gives it several ways to benefit if AI expands flash demand.

The same breadth creates an execution burden. Each market requires different partners, qualification cycles, software, packaging, and support. Developing promising components across categories does not guarantee leadership in any single category.

Samsung also remains active across mobile and enterprise storage while controlling a major DRAM business. SK hynix owns Solidigm, giving it an established enterprise SSD operation. Its strength in HBM provides direct relationships with AI accelerator customers.

Kioxia lacks a comparable HBM franchise. That absence can sharpen its focus on NAND, but it also limits access to the industry’s most profitable AI memory segment. Kioxia must persuade customers that flash deserves a larger share of system investment.

The company’s argument becomes stronger as models and retrieval collections outgrow economical DRAM capacity. It becomes weaker when software can shrink models, improve caching, or avoid frequent storage access.

This is the real competitive line behind the Google News framing. Kioxia is not simply racing Samsung and SK hynix toward a higher number. It is arguing that optimized NAND will become a more important computational resource across AI devices.

What Google News Readers Should Watch Next

Three signals will show whether Kioxia has a competitive AI storage strategy or only a well-timed collection of announcements.

The first signal is production and customer qualification for UFS 5.0. Kioxia expects mass production by the end of 2026, while Samsung targets the fourth quarter. Meeting those schedules would move the contest from samples into supply contracts.

A named smartphone, headset, or edge-device customer would strengthen Kioxia’s position. Independent testing inside that product would matter even more. Readers should look for loading latency, random-read behavior, battery use, and sustained thermal performance.

If Samsung ships broadly while Kioxia remains at the sample stage, the competitive claim weakens. If both reach flagship devices, customers gain a genuine second-source market. Kioxia would not need the highest sequential number to win meaningful business.

The second signal is a complete PCIe 6.0 enterprise platform using Kioxia’s newer BiCS flash. That means more than displaying NAND at a conference. It requires a controller, firmware, form factor, host system, and performance results under realistic workloads.

Sustained mixed-workload throughput will be more informative than a short sequential benchmark. Tail latency, power, cooling, and endurance will reveal whether the product can operate economically in dense AI servers.

Customer qualification would strengthen the argument that faster raw NAND enables useful PCIe 6.0 storage. A prolonged gap between component sampling and finished drives would show that the surrounding system remains the harder problem.

The third signal is how Samsung and SK hynix position flash beside their HBM businesses. SK hynix and SanDisk are already advancing HBF standardization. Samsung can connect mobile storage, enterprise SSDs, processors, and advanced memory within a wider portfolio.

A public accelerator partnership, common HBF specification, or production roadmap would intensify pressure on Kioxia. Conversely, fragmented standards would give Kioxia time to promote conventional SSDs and its own flash-search software.

Readers should also separate corporate measurements from independent evidence. Every vendor has incentives to choose favorable capacities, workloads, and comparison points. Production devices will expose how these technologies behave outside demonstrations.

For developers, the practical question is whether faster storage changes application design. Local models can become larger, while retrieval collections can remain on flash. Software may need new caching, prefetching, and indexing strategies to use that capacity effectively.

Enterprise buyers should watch total system behavior. Faster storage only creates value when accelerators spend less time waiting or when servers need less expensive DRAM. Procurement decisions should follow workload measurements, not interface labels.

The current evidence supports a measured conclusion. Kioxia has credible UFS 5.0 samples, fast next-generation NAND, and a coherent reason for connecting both to AI. It has not established an across-the-board lead over Samsung or SK hynix.

The Google News headline captures the competitive direction but overstates the product-level result. Kioxia is coming for its rivals through specialization, portfolio breadth, and a bet on NAND’s expanding role. The decisive proof will arrive through qualified devices, production schedules, and repeatable system benchmarks.

Watch the first retail hardware closely. Does Kioxia’s storage reduce model-loading delays without increasing heat or battery drain? Does its PCIe 6.0 roadmap produce a qualified enterprise drive that keeps accelerators supplied under sustained workloads? Those results will determine whether AI changes Kioxia’s competitive position or merely gives every flash vendor a new marketing vocabulary.