Hook SanDisk walked into its investor day with a slide deck that looked like a death sentence for HBM. Their HBF (High Bandwidth Flash) solution, they claimed, could match HBM's 12.8 TB/s bandwidth while offering vastly more capacity. The implication: AI inference workloads could ditch expensive DRAM-based HBM for cheaper NAND flash. But Citrini analyst Zephyr tore the comparison apart within hours. The raw data, when you verify it, tells a different story. SanDisk didn't compare HBF to the current or near-future HBM—they cherry-picked a conservative HBM3E configuration and locked it in amber. It's a classic framing tactic, and the crypto-native reader should recognize the pattern: the same kind of parameter manipulation I've seen in DeFi audits where protocols 'prove' their TVL is higher by excluding liquid staking derivatives.
Context The HBM market is a seller's oligopoly. SK Hynix, Samsung, and Micron control the supply, and NVIDIA is the dominant buyer. Prices are high, and capacity is constrained. Enter SanDisk, the NAND flash IDM recently acquired by Western Digital, with a proposal: use 3D NAND in a high-bandwidth package, trade latency for capacity, and undercut HBM on cost per gigabyte. It's a compelling narrative, especially for AI inference where model sizes are ballooning. But the devil is in the parameter choice. SanDisk's HBM baseline was a 192 GB, 12.8 TB/s system using 8 stacks of HBM3E (12-high). That's a 2024 spec. The JEDEC roadmap already shows HBM4E with 16-high stacks reaching 512 GB and 32 TB/s per GPU. By using the older, lower-end metric, SanDisk made HBF look like a bandwidth equal while touting its capacity advantage. Code doesn't lie, but the context around the code (or in this case, the datasheet) can be carefully curated.
Core Let's break down the technical reality. HBF is NAND flash in a 3D stack with TSV-like interconnects. NAND flash has a latency measured in microseconds (μs), while DRAM operates in nanoseconds (ns). That's a 1000x gap. For training workloads, where millions of parameter updates per second require low latency, HBF is a non-starter. For inference, the picture is more nuanced. With quantization to FP4 or FP8, a 480B-parameter model like Qwen3-480B-A35B can fit into 240–480 GB. SanDisk's argument is that HBM cannot reach that capacity cost-effectively, so HBF is the solution. But Zephyr's rebuttal shows that HBM4E at 512 GB already covers that range. The bandwidth difference remains: HBM4E at 32 TB/s vs HBF at 12.8 TB/s. The narrative is a trap; the data is the anchor. SanDisk's presentation implicitly assumed that future AI models would not adopt quantization, or that HBM capacity would stagnate. Both assumptions are false. From my own experience building predictive models for Bitcoin ETF inflows, I learned that to forecast correctly, you must use the latest, not the most convenient, data points. The same applies here.
Contrarian The real story isn't about HBF vs. HBM. It's about the DRAM vs. NAND flash industry war. HBM is a premium product with high margins, controlled by a few players. SanDisk, as a NAND manufacturer, has no seat at that table. So they're building a parallel table: HBF, which uses their existing NAND fabs but requires new high-bandwidth packaging. The market they're targeting isn't training—it's the 'memory pool' layer for AI inference, where large, cold models are stored and paged into GPU memory. This space is currently served by CXL-attached memory and SSDs with smart offloading. HBF is a new contender, but it's not a direct HBM replacement. The controversy around the parameter comparison is a distraction. The real question is: can SanDisk deliver a reliable, high-bandwidth NAND stack at a price point that makes sense for hyperscalers? And can they do it before HBM4E makes the capacity argument moot? The answer is likely no, but the narrative is powerful enough to buy time.
Takeaway Don't trust, verify—always. SanDisk's HBF vs HBM slide is a masterclass in framing, but it falls apart under scrutiny. The crypto community should take note: the same pattern appears in every new L1 whitepaper that claims to 'solve' Ethereum's scalability by comparing against a static version of the chain. The next JEDEC standard for HBM4E will be released within 12 months. If SanDisk cannot match the bandwidth and latency trajectory, HBF will remain a footnote. Watch the packaging partnerships—if SanDisk teams up with TSMC or ASE for advanced interposers, the story changes. Until then, treat HBF as a flash-optimized product, not a HBM killer. The market is a machine for absorbing liquidity, and this debate is just another liquidity event for the semiconductor bull case.
