The press forgot to check the timestamps.
On August 14, a Citrini analyst named Zephyr publicly questioned SanDisk's HBF vs. HBM comparison during an investor day. The numbers looked clean. SanDisk claimed their HBF (High Bandwidth Flash) could match HBM bandwidth at 12.8 TB/s while requiring fewer GPUs for inference. But the ledger tells a different story. The parameters SanDisk selected weren't just outdated—they were strategically frozen in time. HBM3E specifications from 2023 were compared against a speculative HBF product that hasn't shipped. The ledger remembers what the press forgets: benchmark comparisons are only as honest as the parameters you choose to highlight.
Context: The Storage Media Cold War
This isn't just a technical argument. It's a turf war between DRAM and NAND ecosystems. HBM (High Bandwidth Memory) is a JEDEC-standard DRAM stack, built with TSV technology and nanometer-level latency. It's the gold standard for AI training and high-frequency inference. HBF, according to SanDisk's presentation, is a NAND Flash-based alternative using a similar high-bandwidth interface. The core claim: HBF offers larger capacity per dollar with enough bandwidth for inference workloads.
But here's the contextual problem. SanDisk chose to compare their HBF against an HBM configuration with 192 GB total capacity (8 stacks of 24 GB HBM3E 12Hi) and 12.8 TB/s bandwidth. Based on my experience auditing Tether reserves in 2017, I know that parameter selection is a form of storytelling. You can make any product look competitive if you pick the right baseline. SanDisk selected a static HBM specification that is already being superseded by HBM4 roadmaps. The real question isn't whether HBF can match a 2023 HBM benchmark—it's whether it can survive the next 18 months of DRAM evolution.
Core: The On-Chain Evidence Chain
Let's trace the numbers. SanDisk's presentation set both HBM and HBF total bandwidth at 12.8 TB/s, or 1.6 TB/s per stack. This corresponds to a conservative HBM3E performance level. Zephyr's rebuttal proposed a different baseline: 16-layer HBM4E with 8 stacks, yielding 512 GB capacity and approximately 32 TB/s bandwidth. That's 4 TB/s per stack—roughly 3x the bandwidth SanDisk used.
Capacity is the silent variable. SanDisk's demo likely assumed a model like Qwen3-480B-A35B running in bfloat16 precision. At 480 billion parameters, bfloat16 requires roughly 960 GB of memory. HBM's 192 GB can't hold it. HBF's larger capacity becomes the hero. But the narrative shifts when you introduce quantization. FP4 or FP8 precision compresses the same model to 240-480 GB. HBM4E's 512 GB configuration covers that. The capacity advantage evaporates.
Yields are just risk with a prettier name. SanDisk's argument hinges on HBM being capacity-constrained. But the HBM roadmap is accelerating. HBM3E to HBM4 to HBM4E—each generation doubles bandwidth and capacity. HBF, if still based on NAND, faces a persistent latency gap measured in microseconds versus nanoseconds. DRAM's write endurance is orders of magnitude higher. For training workloads, HBF is irrelevant. For inference, it competes in a narrower window.
Floor prices are narratives; volume is truth. The real volume of HBM adoption is in NVIDIA and AMD's data center contracts. HBM supply is locked for years. SanDisk is proposing a new supply chain, a new packaging process, and a new controller architecture. The volume of actual HBF shipments is zero. The narrative is high, but the transaction volume is missing.
Contrarian: Correlation Is Not Causation
Most analysts are framing this as a HBM vs. HBF technology debate. That's a distraction. The real story is SanDisk's strategic positioning against the DRAM oligopoly. SK Hynix, Samsung, and Micron control HBM supply. Prices are high. Allocation is tight. SanDisk, as a NAND IDM, sees an opportunity to capture AI memory spend that would otherwise flow to DRAM vendors.
Trace the coins, not the claims. If HBF ever ships, it won't compete with HBM in training. It will target the inference memory pool and large-capacity caching layers. This is a CXL memory expansion competitor, not a DRAM replacement. The entire presentation was designed to make investors think "HBF is the next HBM," when in reality it's a NAND product with a high-bandwidth interface. The technological distance between NAND and DRAM is not bridged by packaging tricks.
Silence in the blocks speaks volumes. Notice what SanDisk didn't share: latency benchmarks, endurance cycles, power consumption, and production timelines. These are the metrics that matter. If HBF has 10x the latency of HBM, it doesn't matter if the bandwidth matches. The data is silent on the most critical trade-offs.
Efficiency hides the friction points. SanDisk's presentation claims fewer GPUs are needed because HBF offers more capacity per dollar. That logic assumes all GPUs are identical in performance. But if HBF's higher latency forces GPU idle time, the efficiency gain disappears. The friction point is the NAND-to-DRAM gap, and SanDisk is obscuring it with bandwidth parity charts.
Wash trading wears a digital mask. This controversy is a form of marketing wash trading. SanDisk is trading on HBM's reputation by association, while their product is fundamentally different. The comparison framework is designed to mislead, not to inform. Investors should audit the flow, not just the figure.
Takeaway: The Next Week Signal
Watch for HBM4E announcements from SK Hynix and Samsung in the coming months. If they deliver 32 TB/s bandwidth and 512 GB capacity by 2025, SanDisk's HBF narrative collapses. The only credible use case left would be price-sensitive inference markets where latency is less critical. The ledger remembers: bandwidth parity is a snapshot, but the technology roadmap is a movie. SanDisk is showing a still frame. The market will demand the full reel.