SK Hynix and Sandisk May Have Just Solved AI's Biggest Bottleneck — An – Good Machine

TL;DR

SK Hynix and Sandisk published the first open High Bandwidth Flash (HBF) specification through the Open Compute Project on 3 August 2026, targeting up to 512 GB capacity and 3 TB/s bandwidth per stack.
HBF fills the yawning gap between HBM (fast, tiny, expensive) and SSDs (slow, huge, cheap) — a gap that is currently the single largest bottleneck in AI inference at scale.
Google and Tenstorrent are already in the consortium. Samsung is building a rival architecture called zHBM. This is a standards war, not a product launch.
Samples expected H2 2026. Inference devices early 2027. Real revenue is at least a year away. The stock market has not yet priced this in for either SK Hynix or Sandisk.

What Happened

On 3 August 2026, SK Hynix and Sandisk released the first open technical specification for High Bandwidth Flash (HBF) through the Open Compute Project (OCP), coinciding with the opening of the Future of Memory and Storage (FMS) 2026 conference in Santa Clara, California (SK Hynix Newsroom, 3 Aug 2026).

The specification is not a product. It is a blueprint. It defines the host interface, electrical requirements, reliability guidance, packaging recommendations, and software behaviour needed to build HBF-compatible devices. SK Hynix disclosed three scalable bandwidth tiers — Grade 1 through Grade 3 — delivering host interface bandwidth from roughly 400 GB/s up to 3 TB/s. Capacities reach 512 GB per stack using 8-high and 16-high NAND configurations (Hot Hardware, 4 Aug 2026; NERDS.xyz, 3 Aug 2026).

HBF uses the UCIe (Universal Chiplet Interconnect Express) specification, which means HBF die stacks can sit directly on the same silicon interposer or high-density organic substrate as the host GPU or custom ASIC — not across a PCIe bus (TheStreet, 4 Aug 2026). This is the architectural move that makes HBF something genuinely new rather than just a faster SSD.

Alongside the HBF announcement, SK Hynix previewed its tenth-generation V10 375-layer 4D NAND, claiming 2.5× the power efficiency of standard server flash (SK Hynix Newsroom, 3 Aug 2026).

The consortium currently includes Google and Tenstorrent. SK Hynix EVP Kim Chun-sung and VP Kang Uk-song delivered a joint keynote titled "Orchestrating Efficient AI Infrastructure through Tiered Memory in the Era of Agentic AI" on the opening day of FMS 2026. A panel discussion titled "Breaking the Memory Wall with HBF" featuring SK Hynix VP Lim Eui-cheol, Sandisk VP Rajeev Nagabhirava, and Google DeepMind Senior Staff Engineer Xiaoyu Ma is scheduled for 6 August (The Korea Times, 4 Aug 2026).

What It Actually Means

The Memory Wall Is Real, and It Is Expensive

AI inference at scale has a physics problem. The data that large language models need to read — model weights — sits in storage. But storage is slow. The solution today is to load weights into High Bandwidth Memory (HBM), which is fast but eye-wateringly expensive and capacity-constrained. A single H100 GPU has 80 GB of HBM. A frontier model with 1 trillion parameters needs roughly 2 TB just to hold its weights at FP16. That means 25 GPUs before you process a single token.

The gap between HBM and SSDs is where inference costs live. HBF is designed to sit in that gap: faster than SSDs, cheaper and denser than HBM, non-volatile (so it doesn't need constant power), and connected directly to the processor via UCIe rather than through a PCIe bottleneck.

This is not a replacement for HBM. It is a complement. HBM handles the compute-bound hot path. HBF handles the capacity-bound warm path — the weights and embeddings that need to be read frequently but not continuously. The tiered memory architecture SK Hynix is proposing is: HBM for active computation, HBF for nearline model storage, SSD for cold storage.

The Standards War Has Already Started

The specification was published through OCP, which means it is open. Anyone can build to it. That is both the strength and the vulnerability of the strategy.

Samsung is expected to unveil its own next-generation architecture, called zHBM, at the same FMS 2026 conference this week. Unlike HBF, zHBM stacks memory directly on top of the GPU rather than beside it, and it sits outside the SK Hynix–Sandisk consortium entirely (TheStreet, 4 Aug 2026, citing BigGo Finance).

Kioxia, the other major NAND player, has not yet declared a position.

This is the classic standards-war dynamic: SK Hynix and Sandisk are trying to establish HBF as the default through openness and early ecosystem building (Google, Tenstorrent). Samsung is betting on a proprietary integration play with NVIDIA. The winner will be determined not by technical merit alone but by who locks in the largest hyperscaler customers first.

The Timing Is Strategic

SK Hynix formed the HBF consortium with Sandisk only six months ago. Publishing a full OCP specification in that timeframe is aggressive. The company is clearly trying to establish HBF as the de facto standard before Samsung's zHBM gains momentum.

Sandisk's own roadmap targets initial HBF samples for the second half of 2026, with AI inference products following in early 2027 (TheStreet, 4 Aug 2026). That puts real revenue at least a year away. But the specification itself is the product right now — it is a signal to the hyperscaler market that there is an open, consortium-backed alternative to whatever Samsung is cooking.

Hype Deconstruction

This is not shipping hardware. The specification is a blueprint. Samples are expected H2 2026. Inference devices in early 2027. Broader commercial adoption follows after that. The "solved AI's biggest bottleneck" framing is premature — the bottleneck is solved on paper, not in production.

This is not a replacement for HBM. HBF is a new tier between HBM and SSDs. It does not eliminate the need for HBM. It reduces the pressure on HBM capacity, which is valuable, but it is an addition to the memory hierarchy, not a substitution.

The consortium is small. Google and Tenstorrent are credible early partners, but two companies do not make an ecosystem. The consortium needs at least one more major hyperscaler (Microsoft, Amazon, or Meta) and ideally a second NAND manufacturer to reach critical mass.

Samsung is not standing still. The zHBM announcement this week could shift the narrative. If Samsung can demonstrate a working prototype with NVIDIA integration, the HBF consortium's first-mover advantage evaporates.

Stakeholder Landscape

SK Hynix (SKHY): The world's second-largest memory chipmaker and dominant HBM supplier to NVIDIA. HBF extends its memory leadership into a new tier. If the standard is adopted, SK Hynix captures both the HBM and HBF layers of the AI memory stack.

Sandisk (SNDK): Spun off from Western Digital in 2025. Specialises in NAND flash. HBF gives Sandisk a differentiated play in AI infrastructure beyond commodity NAND. The partnership with SK Hynix provides manufacturing scale it would not have alone.

Samsung: The elephant in the room. Samsung is the largest NAND manufacturer globally and has its own HBM business. If zHBM gains traction with NVIDIA, Samsung could render the HBF consortium irrelevant. If HBF wins, Samsung faces a competitor-controlled standard in a market it expects to dominate.

NVIDIA: The silent kingmaker. NVIDIA's GPU architecture determines what memory standards hyperscalers adopt. If NVIDIA integrates HBF support into its next-generation platform, the standard is validated. If it goes with Samsung's zHBM, HBF becomes a niche.

Google (GOOGL): Already in the HBF consortium. Google's TPU architecture gives it an independent path to adopt HBF regardless of what NVIDIA does. Google's participation signals that at least one major AI infrastructure buyer is serious about the technology.

Hyperscalers (Microsoft, Amazon, Meta): The ultimate customers. Their procurement decisions will determine which standard wins. They benefit from the competition either way — a standards war drives down memory costs.

AI model developers (OpenAI, Anthropic, etc.): Indirect beneficiaries. Cheaper inference infrastructure means cheaper API calls and the economic viability of larger models. But they have no direct influence on the standards outcome.

Cross-Layer Implications

Security: A new memory tier introduces a new attack surface. HBF sits on the same interposer as the GPU, which means a compromised HBF controller could potentially snoop on model weights in transit. The OCP specification includes reliability guidance but the security model for chiplet-level interconnects is immature.

Supply chain: HBF uses NAND, not DRAM. NAND manufacturing capacity is less concentrated than HBM capacity (which is effectively an SK Hynix–Samsung duopoly). If HBF takes off, it could reduce the AI industry's dependence on the HBM supply chain — which is currently the single largest hardware constraint on AI scaling.

Geopolitics: SK Hynix is a South Korean company. Sandisk is American. The HBF consortium is structured as a US–Korea partnership, which matters for export control purposes. If HBF becomes critical AI infrastructure, it will almost certainly fall under US export controls on advanced semiconductor technology — potentially limiting its availability to Chinese AI companies.

Energy: SK Hynix claims the V10 375-layer NAND used in HBF is 2.5× more power-efficient than standard server flash. At hyperscale, the power savings from replacing HBM capacity with HBF capacity could be material — potentially tens of megawatts per data centre. This is the underappreciated story: HBF is not just a performance play, it is an energy play.

What This Means for You

For AI infrastructure engineers: Watch the FMS 2026 panel on 6 August. The Google DeepMind participation is the signal to track — if Google commits to HBF in its TPU roadmap, the standard has a real chance. Start modelling tiered memory architectures in your inference pipelines. The HBM → HBF → SSD hierarchy will require new data placement strategies.

For investors: Neither SK Hynix nor Sandisk has seen significant stock movement on this announcement. The market is correctly treating it as a specification, not a product. Revenue impact is at least 12–18 months out. The real trade is on the standards outcome: if HBF wins, SK Hynix and Sandisk benefit disproportionately. If zHBM wins, Samsung consolidates its position. Position sizing should reflect the binary nature of standards wars.

For AI application developers: This is infrastructure plumbing. It will matter to you when inference costs drop, which will happen if HBF or zHBM succeeds. In the near term, nothing changes. In the medium term, cheaper inference unlocks use cases that are currently economically unviable — think always-on AI agents, real-time video understanding, and large-scale model ensembles.

For everyone else: This is a story about the physical limits of AI scaling. The fact that a new memory tier is needed at all tells you something about how fast the industry is hitting walls. The memory wall is one of several physical constraints — alongside power, cooling, and chip fabrication capacity — that will shape what AI can and cannot do over the next five years.

Uncertainty Ledger

Will Samsung's zHBM announcement this week change the competitive landscape? If zHBM demonstrates a working prototype with NVIDIA, the HBF consortium's narrative weakens considerably.
Will additional hyperscalers join the HBF consortium? Microsoft, Amazon, or Meta joining would be a strong signal. Their absence is notable.
Can HBF meet its sampling timeline? H2 2026 is aggressive for a new memory standard. Delays are common in memory technology transitions.
What is the actual performance delta in production? The specification numbers (3 TB/s, 512 GB) are theoretical maximums. Real-world performance in inference workloads is unknown.
Will US export controls apply to HBF? If HBF is classified as advanced semiconductor technology, it could be restricted — which would limit the addressable market but also create a moat for the consortium members.

Bottom Line

SK Hynix and Sandisk have published a credible, technically detailed specification for a new memory tier that addresses a genuine and growing bottleneck in AI inference. The technology is real. The consortium is small but serious. The timeline is aggressive but plausible. The real story is not the specification — it is the standards war with Samsung that is now openly underway. Whoever wins determines the memory architecture of the next generation of AI infrastructure. The fight starts this week in Santa Clara.

Sources:

SK Hynix Newsroom, "SK hynix Unveils First HBF Standard Specifications with Sandisk," 3 August 2026 (Tier 1 — official corporate PR)
TheStreet, "SK hynix, Sandisk race to fix AI's next chip bottleneck," 4 August 2026 (Tier 2)
Hot Hardware, "SK Hynix And SanDisk Unleash High-Bandwidth Flash To Fix AI Bottlenecks," 4 August 2026 (Tier 2)
The Korea Times, "SK hynix, Sandisk unveil standards for high-bandwidth flash memory," 4 August 2026 (Tier 2)
NERDS.xyz, "Sandisk and SK hynix want HBF to break the AI memory bottleneck," 3 August 2026 (Tier 3)
24/7 Wall St., "SK Hynix and Sandisk May Have Just Solved AI's Biggest Bottleneck," 4 August 2026 (Tier 3)
Seeking Alpha, "SK hynix and Sandisk unveil first high bandwidth flash standards," 4 August 2026 (Tier 3)
WCCFTech, "SK hynix, In Collaboration With SanDisk, Unveils The New High Bandwidth Flash (HBF) Standard," 4 August 2026 (Tier 3)