Hook
On August 14, a Citrini analyst named Zephyr dropped a grenade into SanDisk's investor day narrative. The target: SanDisk's comparison of its proposed High Bandwidth Flash (HBF) against HBM. The accusation was simple—SanDisk cherry-picked HBM parameters to make HBF look like a viable alternative. But beneath the technical squabble lies a deeper truth about how memory vendors are positioning for the AI inference boom. And as someone who spent years modeling cross-border payment liquidity, I know a benchmark manipulation when I see one.
Context
SanDisk, a NAND flash IDM, unveiled HBF as a high-bandwidth flash memory stack designed to compete with HBM in AI workloads. The demo showed HBF and HBM both delivering 12.8 TB/s total bandwidth—1.6 TB/s per stack—with HBM at 192 GB capacity (8×24 GB HBM3E 12Hi) and HBF presumably offering much higher capacity at lower cost. The argument: HBF can serve more AI models per GPU, reducing the number of GPUs needed for inference.
Zephyr countered that SanDisk used a conservative HBM baseline. The analyst proposed a 16-layer HBM4E configuration with 8 stacks delivering 512 GB and 32 TB/s bandwidth—roughly 3× the bandwidth per stack. At bfloat16, a 480B-parameter MoE model like Qwen3-480B-A35B would require 240–480 GB depending on quantization. Under FP4/FP8, the capacity requirement shrinks further, meaning HBM4E could cover the same use case that HBF claims as its selling point.
Core Insight: The Technical Feasibility Check
From my experience building Python simulations for payment rail efficiency, I learned that the frame of comparison determines the conclusion. SanDisk's 12.8 TB/s bandwidth is real for HBM3E, but it’s already outdated. The HBM roadmap is accelerating: HBM3E → HBM4 → HBM4E, with each generation doubling bandwidth and capacity. HBM4E is expected to reach 4 TB/s per stack, making SanDisk's static comparison a snapshot of a moving target.
What SanDisk didn't show is the latency and endurance gap. HBM uses DRAM with nanosecond latency and nearly unlimited write cycles. HBF, based on NAND flash, operates at microsecond latency and has limited write endurance. In AI training, where every nanosecond of stall costs GPU cycles, HBF is a non-starter. But for inference—especially batch inference with large models—the latency penalty can be absorbed by prefetching and caching.
The real hidden parameter is quantization. SanDisk’s demo likely assumed bfloat16 to maximize the capacity advantage of HBF. Under FP4, the same model fits in 240 GB, which HBM4E can handle. That closes the gap. Zephyr’s critique is valid: SanDisk chose a conservative HBM spec to amplify HBF's relative strength. This is a classic benchmark framing tactic, and I’ve seen it in every tech cycle from payment rails to blockchain consensus.
Contrarian Angle: HBF Isn't Trying to Replace HBM
Here’s where the market narrative diverges from the technical debate. SanDisk's HBF is not a direct HBM competitor. It’s a large-capacity memory pool for the inference tier, competing with CXL memory expansion and SSD-based caching. The real war is between DRAM and NAND capital. HBM prices are exorbitant and supply is locked by NVIDIA and AMD. SanDisk sees an opening for a “good enough” high-bandwidth flash that can serve models requiring hundreds of gigabytes but not nanosecond latency.
Most analysts miss this: HBF’s true target is not the training rack but the inference server. In decentralized AI inference networks—which crypto projects are building—cost per gigabyte matters more than absolute latency. A flash-based memory pool can reduce the capex for running large models, enabling more nodes on the network. That’s where HBF could disrupt the status quo, not by matching HBM’s performance but by offering a different point on the cost-latency curve.
Takeaway: The Memory Hierarchy Is Shifting for AI
SanDisk’s HBF controversy is a microcosm of a larger shift: the memory hierarchy is being rearchitected for AI workloads. HBM will dominate training and latency-sensitive inference, but HBF may carve out a niche in high-capacity, cost-sensitive scenarios. For crypto-focused AI infrastructure, the implication is clear: watch the memory cost curve, because it determines how many agents can run autonomously. The market will decide which layer of the memory hierarchy gets the capital. And right now, the smart money is betting on both DRAM and NAND finding their place in the AI stack.
