SanDisk just dropped an architecture that turns NAND flash into AI memory. Market cap of HBM? $160B. HBF aims to undercut it by 50%.
Hook: The pitch is simple—take cheap NAND, stack it high, call it High Bandwidth Flash. Conventional wisdom says AI memory requires DRAM. HBF says: not for inference. And that's where the alpha lies.
Context HBM dominates AI training—SK Hynix, Samsung, Micron have locked the supply chain with EUV and CoWoS. But inference is different. Inference workloads need massive capacity to hold large models, not ultra-low latency. A single LLM inference server can require 1TB+ of memory. HBM at $20/GB kills the economics. NAND at $0.10/GB changes the game.
SanDisk knows this. After splitting from WD, they needed a new narrative. HBF is that narrative—a storage architecture repurposed as memory. But the devil is in the physics.
Core: The Architecture Trade-Off
HBF uses 3D NAND dies stacked with TSV interconnects, similar to HBM but without the DRAM. The bandwidth targets are lower—think 100-200 GB/s per stack vs HBM3e's 1 TB/s. But the capacity per stack can exceed 64GB, far beyond HBM's 16-24GB. For inference, you don't need nanosecond latency; you need to fit the model in memory.
Key technical findings from the layout:
- NAND latency is microseconds. DRAM is nanoseconds. 1000x slower. But inference batch processing can tolerate this if the model is cached properly.
- Cost per GB is 5-10x less than HBM. Even with packaging overhead, HBF can undercut HBM by 50%+.
- No EUV required. NAND fabs use DUV, which faces fewer export controls. This is a geopolitical hedge—HBF can be manufactured in Japan (Kioxia JV) without triggering US chip bans.
But here's the hidden signal: SanDisk is not trying to beat HBM on performance. They are creating a new memory tier. The compute trend is moving toward disaggregated memory pools via CXL. HBF could slot into that pool as a cost-effective, high-capacity layer. Alpha isn't found in memes; it's buried in silicon.
Contrarian: The Market Is Missing the Supply Chain Crack
Everyone is obsessed with HBM4's bandwidth. The real risk? HBM supply is constrained by CoWoS capacity and EUV allocation. SK Hynix is spending $15B+ on new HBM fabs. That capex drives depreciation, which raises prices.
SanDisk's HBF flips the script. They don't need new fabs—they can repurpose existing NAND capacity from the Kioxia JV. The NAND industry is still recovering from the 2023 glut. HBF gives that idle capacity a higher-margin use case. Yields are the reward for paranoia.
But the contrarian angle goes deeper. HBF is a statement about the future of AI inference markets. The training market is saturated by hyperscalers. Inference is fragmented—edge AI, on-device LLMs, autonomous agents. Traditional HBM is overkill for these use cases. SanDisk is positioning for a world where AI models are deployed everywhere, not just in data centers.
However, the path is treacherous. The ecosystem is nascent. No controller, no driver, no OS support. Cloud providers are locked into HBM supply chains. And if Samsung or SK Hynix launch a "HBM Lite" product at similar cost, HBF's window closes.
Takeaway: Watch the Signals
HBF is not a product yet. It's a bet. The key metric to track isn't bandwidth—it's customer adoption. If AWS or Meta announces a joint design win, then the narrative changes. Until then, treat this as a capital allocation signal: SanDisk is betting its future on AI inference memory. The question is whether the market will bet alongside them.
Actionable levels:
- If HBF secures a JEDEC standard or CXL integration, it's a buy signal.
- If no major cloud partner within 12 months, the thesis fails.
Audit the code, ignore the influencer. In this case, the code is the silicon. The influencer is the hype. I've seen this movie before—in 2017 with ICO arbitrage, in 2020 with DeFi audits, in 2022 with Terra shorts. The contrarian trade is always the one that builds on fundamentals, not narratives. HBF has fundamentals. Now it needs execution.