HBF Enters Standardization; How Much It Can Cut Inference Costs Remains Unproven

The Standard Establishes a Starting Point for Collaboration

Sandisk announced on August 3, 2026 that it and SK hynix had released the first HBF technical specifications through OCP, with Google and Tenstorrent participating in the alliance and technical validation. The specifications cover interfaces, electrical requirements, packaging, reliability and software read/write guidelines, and allow HBF to coexist with HBM. This gives participants a shared set of technical conventions on which to base design and validation. My assessment is that publishing the standard reduces uncertainty in collaboration, but member participation demonstrates technical commitment. Whether customers will buy, and when volume production will generate revenue, still require separate evidence.

Evidence: [1]

The Value of an Additional Tier Depends on How Data Is Used

On August 4, 2026, SK hynix positioned HBF between HBM and SSDs, using NAND to expand capacity. Its published specifications include 8-layer and 16-layer stacks, capacity of up to 512 GB, three bandwidth tiers spanning approximately 0.4 to 3.0 TB/s, and a UCIe connection interface. This suggests that HBF's technical direction is to add a tier that the system can manage. Whether it delivers economic benefits depends on whether actual tasks can use the added capacity effectively and whether their data access patterns suit this tier. These figures are vendor-published specifications. Capacity and bandwidth alone do not address random-access latency, software scheduling overhead or actual service costs.

Evidence: [2]

Kioxia Offers Another Interface Path

Kioxia's GP1, announced on August 4, 2026, uses PCIe 6.0, NVMe and second-generation XL-FLASH, with up to 10 million IOPS for 512-byte random reads. Evaluation samples are scheduled for selected customers by the end of 2026, and sample specifications may differ from those of production products. Compared with HBF's announced tier and connection approach, GP1 presents a path for advancing high-performance flash through an SSD interface. Both merit comparison under the same research question: how can flash serve AI tasks more effectively? But HBF's bandwidth specifications and GP1's random-read metrics measure different things and cannot be used to rank their performance directly.

Evidence: [3] [2]

Demonstration Targets Rise, but Commercial Stages Still Need to Be Distinguished

On October 8, 2026, Kioxia further announced plans to present an enhanced GP-series technology demonstration using third-generation XL-FLASH at the October 12–15 OCP summit, targeting more than 20 million IOPS for 512-byte random reads. This adds a performance target worth validating, but does not directly establish that GP1 evaluation samples or future production products achieve the same performance. The product descriptions and technology stages in these announcements need to remain distinct. What can currently be confirmed is that the company announced a demonstration plan. Disclosed results are still insufficient to assess the on-site performance, test conditions or effectiveness in actual AI tasks.

Evidence: [4] [3]

Cost Judgments Should Wait for System Results

Taken together, these announcements lead me to conclude that high-performance flash now has a concrete basis for validation within AI memory and storage architectures, but it is too early to specify a definite cost reduction. This judgment rests on one premise: an additional tier must deliver enough benefit in the target tasks to offset the costs of integrating and using it. The HBF standard and GP-series plans have not yet provided comparable results for complete services. New evidence that could change this assessment should include clearly defined workloads, comparison configurations, service latency, sustained throughput and corresponding costs. Customer orders and volume production deliveries would help assess commercial progress. Even if a test shows an advantage, its conditions of applicability should be established before discussing whether the result can be generalized.

Evidence: [1] [2] [3] [4]

What to watch next

• Whether HBF system tests are published with clearly defined workloads, comparison configurations, service latency and sustained throughput. Whether the results support a conclusion that costs can be reduced.
• Whether Kioxia discloses the enhanced GP-series demonstration results and test conditions, and explains how they relate to GP1 evaluation samples and production specifications.
• Whether vendors disclose customer orders, volume production deliveries and cost information, allowing technical validation to be linked more closely to commercial results.