SK Hynix and SanDisk Release First HBF Standard Specification; Google Joins the Ecosystem
Alina Collins
SK hynix and Western Digital unveiled the first standard specification for High-Bandwidth Flash (HBF) at FMS 2026 — just six months after the alliance formed. Google is in, but Nvidia remains on the sidelines — making this a standards race where the biggest swing factor is who shows up.
What is HBF, and why does AI need it now?
HBF — High-Bandwidth Flash, a new storage tier sitting between memory and SSDs — tackles two AI-inference pain points at once: not enough bandwidth, not enough capacity.
This means → today's HBM (high-bandwidth memory, the ultra-fast memory stacked right next to the chip) is fast but small; SSDs are large but slow. HBF is designed to fill that gap.
SK hynix defines it as a key building block of a "tiered memory" architecture. In plain terms = future AI servers won't have to choose between "fast" and "big" — they get a middle layer that's both.
What's in the spec — the key numbers?
Capacity: supports 8-layer and 16-layer NAND die stacking, scaling up to 512 GB.
Bandwidth: three performance grades, ranging from roughly 0.4 TB/s to 3.0 TB/s — This means → the top grade approaches current mainstream HBM speeds, but at far greater capacity.
The interconnect standard is UCIe — Universal Chiplet Interconnect Express, an open interface that lets different chips plug into each other. In plain terms = whether your processor is a GPU, CPU, or something else, HBF can connect to it — no vendor lock-in.
Google is in, Nvidia is watching — why ecosystem alignment matters most?
Google and AI-chip startup Tenstorrent have joined the HBF alliance. The spec is published through the Open Compute Project (OCP), open to the entire industry.
Citrini Research analyst Jukan has noted: Google is the most active backer of HBF adoption, while Nvidia sees existing eSSD as sufficient and shows limited interest.
This means → whether Nvidia ultimately adopts HBF is the single most important validation point for this technology's chances of replicating HBM's dominance in AI-inference hardware. Without Nvidia, HBF risks staying a "plan B"; with Nvidia, it becomes the standard.
From founding to first spec in six months — what does the speed tell us?
The HBF alliance was formally established in February this year. The first standard spec landed on August 4 — far faster than most industry standardization cycles.
This reflects how urgently AI inference demands more storage bandwidth — the industry can't afford to wait for a slow standards process.
SK hynix EVP Kim Chun-sung said HBF will break the boundary between memory and storage, enabling a new architecture that improves overall system efficiency.
One more headline — 10th-gen 375-layer 4D NAND makes its debut?
SK hynix showcased its in-development 10th-generation (V10) 375-layer 4D NAND wafer and products at the FMS 2026 booth.
Performance per watt improves 2.5× over the prior generation, optimized for data-center and AI-infrastructure workloads.
SK hynix plans to begin mass production in early next year. In plain terms = this isn't a lab demo — it's already on the production timeline, aimed at cementing SK hynix's lead in the AI NAND market.
Content is for reference only, not financial advice.