SanDisk introduced High Bandwidth Flash (HBF), a memory module that stacks NAND flash dies, at its Investor Day in Focus 2026 and FMS 2026. The company positioned HBF as a potential successor to High Bandwidth Memory (HBM) for large-scale AI inference, citing the inherent capacity advantage of NAND flash, which offers 10 to 100 times the memory per silicon die compared to DRAM. This development addresses growing concerns that HBM capacity may be insufficient for the memory demands of widespread AI inference systems, where response time and data residency are critical.
HBF 4TB vs HBM 192GB per GPU
In a comparative emulation of Alibaba’s Qwen3-480B-A35B model, SanDisk demonstrated that an HBF-equipped GPU package with 4TB of memory per GPU achieved stable inference performance nearly equivalent to an HBM-equipped package with 192GB per GPU. Crucially, the HBF configuration required only four GPUs to match the performance of an eight-GPU HBM configuration, effectively halving the GPU count for the same output throughput. While the HBF package offered 16TB of total memory compared to 1.5TB for the HBM setup, SanDisk noted that the cost differential between the two configurations cannot be definitively stated without further market data.
NAND core die replaces DRAM in HBF
The structural innovation of HBF replaces the DRAM core die of traditional HBM with NAND flash, maintaining a similar stacking architecture. SanDisk stated that NAND flash read access times are close to DRAM and that read throughput can be comparable to HBM, making it suitable for workloads where capacity is the primary bottleneck. The company first published the HBF concept in February 2025 and announced a collaboration with SK hynix in August 2025 to develop and define specifications for the technology, signaling a potential shift in the memory hierarchy for AI accelerators.
Hybrid HBF and HBM memory allocation
SanDisk proposed several integration strategies for HBF, including replacing all HBM modules in an XPU package to increase capacity by 8 to 16 times, or using a hybrid approach where half the memory is allocated to HBF for persistent data and the other half to HBM for frequently changing data. This flexibility allows system architects to optimize for either massive data residency or high-speed access, depending on the specific stage of AI inference processing. The company suggested that HBF could serve as a high-performance memory extension tier, reducing reliance on costly in-node HBM or DRAM expansion for large datasets (Kioxia).
HBM shortage drives price volatility
The introduction of HBF occurs against a backdrop of escalating HBM supply constraints and price volatility. Reports indicate that Chinese AI chip manufacturers are paying multiples of international prices for advanced HBM due to US export controls, with Huawei’s Ascend 950DT chip price increasing by up to 50% in two months. This cost pressure highlights the urgent need for alternative memory solutions that can provide high capacity without the premium associated with scarce HBM supply, potentially shifting the focus from pure bandwidth to capacity-per-dollar for specific inference workloads.
Kioxia GP1 SSD for GPU direct access
Kioxia’s recent recognition of its GP1 PCIe 6.0 SSD as ‘Best of Show’ at FMS 2026 further underscores the industry’s move toward flash-based memory extensions for AI infrastructure. The GP1 SSD, optimized for GPU direct access, delivers up to 10 million random read IOPS, enabling flash to serve as a high-performance tier that improves GPU utilization. Together with SanDisk’s HBF, these developments suggest a broader trend where NAND flash is increasingly integrated into the core memory hierarchy of AI systems, challenging the traditional dominance of DRAM-based HBM in high-performance computing.