Kioxia and Sandisk Split on High-Bandwidth Flash Strategy

Sandisk taped out its first High-Bandwidth Flash memory die in FQ1, while joint-venture partner Kioxia has declined to join the SK Hynix-led HBF consortium and is building an alternative architecture based on serial, non-stacked flash modules. The two companies share NAND fabs. They no longer share a roadmap.

The split matters commercially. Sandisk owns no independent wafer fabs, so its only fab access today runs through the Kioxia joint venture. HBF dies cannot come from standard NAND wafers, and the JV shows no sign of producing HBF wafers for Sandisk alone.

What is High-Bandwidth Flash?

HBF mirrors the High-Bandwidth Memory concept. Multiple layers of NAND connect through Through-Silicon Vias to a base interposer logic unit, which links the stack to a host GPU in parallel rather than serially. Known-good dies are stacked vertically, typically 8-high or 16-high in the first specification, and interconnected with TSVs and microbumps.

Standard NAND dies cannot serve as-is. The stack requires dies designed from the start for TSV and microbump connectivity, for three reasons:

• TSVs must pass vertically through each die, demanding reserved keep-out zones, extra process steps and additional die area that conventional SSD-packaged 3D NAND dies lack.
• Lower dies in the stack need TSV pathways connecting through to the base logic die and interposer.
• HBF dies require far higher internal parallelism, with more independent planes and sub-arrays, to hit bandwidth targets through concurrent reads.

How far has the HBF spec progressed?

Sandisk and SK Hynix launched the HBF Spec Standardization Consortium and a dedicated workstream under the Open Compute Project in February 2026. The first open technical specification, released via OCP in early August 2026, covers NAND die stacks at 8-high and 16-high up to 512 GB, bandwidth grades from roughly 0.4 to 3.0 TB/s, the UCIe interconnect, packaging, reliability and software guidance. Google and Tenstorrent joined the consortium afterward.

Professor Kim Jung-ho of KAIST, the technology's most prominent advocate, said in January: "Samsung Electronics and Sandisk plan to integrate HBF into products from Nvidia, AMD, and Google by late next year or early 2028."

Sandisk confirmed progress in October: "HBF reached a milestone in FQ1 with tape-out of our first HBF memory die; samples of our first HBF-based products remain on track for next calendar year."

What is Kioxia doing instead?

Kioxia has staked out its own path across flash and SSD development:

• A GP Series SSD using generation 3 XL-FLASH, targeting Optane-class performance with more than 20 million IOPS on 512-byte random reads.
• Liquid-cooled NX1 Series E1.S NVMe SSDs for data centers, plus a liquid-cooled version of its PCIe Gen 5 CM9 datacenter SSD.
• Fast NAND as a DRAM alternative, as reported by Japan's Nikkei, using a CXL module with memory access across a PCIe link. Kioxia claims the combination of high-speed NAND and a CXL controller cuts data-read latency to less than one-tenth of conventional NAND products.
• A prototype high-bandwidth flash module shown in August 2025 for edge servers, not GPU servers, offering 5 TB capacity and 64 GB/s bandwidth, built from daisy-chained "beads" of flash rather than stacked layers, with 128 Gbps high-speed transceiver technology and differential serial signalling instead of the parallel approach SK Hynix and Sandisk favor.

The architectural divide is clean: Kioxia is considering serial connectivity without layer stacking. Sandisk and SK Hynix favor stacked layers with parallel TSV connections.

Where would Sandisk source HBF wafers?

SK Group Chairman Chey Tae-won suggested in September that a partnership between SK Hynix and Kioxia could be interesting. SK Hynix already holds Kioxia convertible bonds that would convert into a 14.19 percent voting stake, the largest single holding, and Kioxia uses SK Hynix DRAM in some SSDs.

Kioxia CEO Hiroo Ota rejected the idea publicly, telling The Japan Times: "We can't just say, 'Well then, let's make it three companies' … We have no idea what prompted (Chey) to say what he did." Such open disagreement between a Korean chaebol chairman and a Japanese CEO is rare; disputes at this level usually stay private.

A TrendForce document from April 2026, citing ETNews, reported that Sandisk has begun engaging materials, components and equipment partners to build an HBF prototype production line. The report said prototypes are targeted for the second half of the year, with Japan a leading candidate for the site, pilot operation around year-end and commercialization targeted for 2027.

That points to a possible scenario in which Sandisk owns and operates its own NAND fab for HBF, separate from the Kioxia JV, a move that would add strain to a partnership already pulling in two architectural directions. Chip Dispatch has asked Sandisk whether a separate HBF fab is under consideration.