The press release landed with the usual choreography: a bold claim about a new memory architecture named HBF (High Bandwidth Flash) that would “redefine AI memory.” The hype cycle has already begun. But the ledger of engineering reality does not lie—it only waits to be read. And the reading is far from favorable.
SanDisk, freshly split from Western Digital, needs a narrative. HBF is that narrative. But a narrative is not a product. The architecture is real in concept, but the data sheet is missing. No latency figures. No bandwidth numbers. No confirmed timeline. The only certainty is the absence of certainty.
Context: The HBM Monopoly and the Desperate Search for an Alternative
For the past two years, the AI memory market has been dominated by HBM (High Bandwidth Memory). SK Hynix, Samsung, and Micron have locked down the supply chain, with HBM3e becoming the standard for AI accelerators. The barriers to entry are astronomical: EUV lithography, advanced packaging (CoWoS), and years of process optimization. SanDisk, a NAND flash manufacturer, is not a player in that game. Its NAND market share sits around 15%, and its HBM share is zero.
Enter HBF: a flash-based memory stack that aims to offer higher capacity at lower cost than HBM. The pitch is simple—AI inference workloads need massive memory capacity to hold large models, but they can tolerate higher latency than training. HBF targets that niche. It is a classic case of a company trying to turn a weakness (no DRAM capability) into a strength (NAND is cheap and abundant). But as I have learned from auditing smart contracts for years, a clever workaround is not a solution until it is proven in production.
Core: A Systematic Teardown of HBF's Claims
Let me be clear: I am not dismissing HBF outright. I am demanding evidence. Based on my experience reverse-engineering the EtherDelta contracts and later modeling the Terra collapse, I have learned to treat architectural promises as hypotheses until the data arrives. Here is what the current announcement lacks and why that matters.
1. The Missing Metric: Latency
HBF is built on NAND flash. NAND read latency is in the microsecond range (10–100 µs). DRAM latency is in the nanosecond range (50–100 ns). That is a factor of 100 to 1,000 difference. For AI inference, the model weights must be fetched from memory for every token generated. If the memory is slow, the inference throughput collapses. SanDisk has not disclosed HBF's actual latency. The claim that “flash can be fast enough for inference” is unsubstantiated. In my forensic work, I have seen countless projects fail because they assumed a physical limitation could be papered over by software. Physics is not negotiable.
2. The Bandwidth Gap
HBM achieves bandwidth of 1–2 TB/s per stack. HBF, using NAND, will likely struggle to reach even 100 GB/s. The architecture may use TSV (Through-Silicon Via) and stacking, but the fundamental I/O speed of NAND is orders of magnitude lower. The blog post mentions “high bandwidth” but does not define it. Without a number, it is a marketing term, not an engineering specification.
3. The Endurance Problem
AI inference involves repeated reads of the same model weights. Flash memory has limited write endurance, but reads are generally safe. However, if the system requires frequent updates of weights (e.g., for fine-tuning), the endurance becomes a concern. SanDisk has not addressed this. The ledger of endurance does not lie: NAND cells wear out after 10,000 to 100,000 program/erase cycles. DRAM is not limited by wear. For inference-only workloads, this might be acceptable. But it is a constraint that must be managed.
4. The Ecosystem Gap
HBM is a standard backed by JEDEC, with mature controller designs, drivers, and integration into GPU packages. HBF is a proprietary architecture. To become a viable alternative, it needs support from server OEMs, OS vendors, and AI framework developers. SanDisk has no such ecosystem. The closest analogy is the failed attempt to replace DRAM with NVDIMM or Intel's Optane. Both were technically impressive but died due to lack of ecosystem adoption. HBF faces the same uphill battle.
5. The Supply Chain Disconnect
SanDisk's NAND production is tied to its joint venture with Kioxia in Japan. The company cannot independently scale capacity. For HBF to reach volume production, it needs to allocate NAND wafers from that JV—wafers that are currently used for SSDs and other products. The opportunity cost is real. And the joint venture governance introduces a dependency that could slow down decision-making. In my analysis of supply chain vulnerabilities in crypto mining, I have seen how such dependencies can become bottlenecks.
6. The Financial Math
HBM manufacturing yields are still ramping, but the margins are high (SK Hynix reports HBM margins above 50%). HBF, if it ever reaches production, will likely be sold at a lower price point to compensate for worse performance. The economics are not in its favor. The analysis in the original report estimates that HBF's gross margins will be far below HBM, and the initial capital expenditure for packaging lines could be significant. The math does not support a rapid return on investment.
Contrarian: What the Bulls Got Right
I am not here to be a pure cynic. The contrarian view deserves a fair hearing. The bulls argue that HBF is not meant to compete with HBM in training; it is for inference, where the market is growing at 70% CAGR. They point to the cost advantage of NAND: per gigabyte, NAND is roughly 10x cheaper than DRAM. If HBF can achieve even 20% of HBM's bandwidth at 30% of the cost, it could be a compelling option for inference servers that need to load large models (e.g., 100B+ parameters) without costing a fortune.
Furthermore, the geopolitical angle is real. The US export controls on HBM to China have created a vacuum. HBF, because it is based on NAND (which is less restricted), could be a legal alternative for Chinese AI companies. The original report's hidden information suggests this is a deliberate strategic move. I have seen similar patterns in the crypto world, where projects built in jurisdictions with lighter regulation gained a foothold. But the compliance risk remains high, and the potential for future restrictions cannot be ignored.
Finally, SanDisk's engineering team has deep expertise in NAND controllers and firmware. They have been building high-performance storage for decades. If anyone can squeeze performance out of NAND, it is them. The architecture is not a pipe dream—it is a plausible engineering challenge. But plausible is not the same as probable.
Takeaway: The Ledger Does Not Lie, It Only Waits to Be Benchmarked
HBF is a strategic narrative designed to boost SanDisk's valuation and give it a distinct identity post-WD split. The market will reward it with a temporary premium. But the real test will come when the first benchmarks are published. If the latency is above 1 microsecond, the bandwidth below 100 GB/s, and the endurance below 100,000 cycles, the narrative will collapse. The ledger of engineering reality does not lie.
I have seen this movie before. In 2018, a dozen projects promised to replace Ethereum with faster, cheaper chains. Most are dead. In 2020, HBM was dismissed as too expensive—until it became the only option. The lesson is that architectural innovation is not enough; you need real-world validation. HBF has none. Track the benchmarks. Ignore the press releases. The only thing that matters is the data.
Until then, the appropriate response is not excitement. It is skepticism. The ledger does not lie, it only waits to be read. And I am still waiting.