The market is obsessed with HBM supply constraints. Every analyst tracks SK Hynix’s TSV capacity, every trader watches Samsung’s CoWoS allocation. But while the crowd watches the flood, a quieter current is forming. SanDisk, freshly split from Western Digital, just unveiled HBF—High Bandwidth Flash. It’s a NAND-based memory architecture aimed directly at the AI inference gap. And if you only see a cheaper HBM, you’re missing the macro signal.
Context: The Architecture of Escape
HBF is not a new node. It’s a structural redefinition of what memory can be. Instead of stacking DRAM dies with advanced packaging (HBM’s path), SanDisk stacks NAND flash dies using TSV and high-bandwidth interconnects, then connects them via a controller optimized for AI workloads. The result: a memory solution with significantly lower cost per gigabyte, but with higher latency (microseconds vs. nanoseconds for DRAM) and lower bandwidth. This is not a training memory. It’s an inference memory—designed to hold the massive model parameter sets that LLMs require during inference, where latency tolerance is higher and cost sensitivity is acute.
Crucially, HBF sidesteps the entire HBM supply chain bottleneck. HBM fabrication requires EUV lithography and advanced packaging equipment (CoWoS, TSV) that are both capital-intensive and geopolitically sensitive. NAND flash, by contrast, uses mature DUV processes and is not subject to the same export controls. SanDisk’s choice is a deliberate hedge: they can’t compete in the HBM capacity race (that’s SK Hynix and Samsung’s game), so they’re creating a new category.
Core: The Macro Logic of the NAND Route
Let’s dissect the structural incentives. First, the AI inference market is exploding. By 2028, inference workloads are projected to grow at over 70% CAGR, driven by deployed models, RAG systems, and edge AI. The bottleneck is not compute—it’s memory capacity and cost. A single 70B parameter model requires ~140GB of memory just for weights. With HBM3e costing roughly $15–20 per GB, a 16-GPU server can burn $40,000+ on memory alone. HBF, if it achieves a 30–50% cost reduction, could cut that bill by half.
Second, the geopolitical angle. The US has already tightened HBM exports to China. Any memory with bandwidth above a certain threshold is now restricted. HBF, by using NAND, operates in a regulatory gray zone. It’s not explicitly covered by current controls, but if it becomes widely adopted for AI inference, it will be. SanDisk’s timing—announcing HBF just as HBM export rules tighten—is not coincidental. It’s a structural hedge against supply chain weaponization.
Third, the NAND industry is emerging from a deep cyclical trough. After 2023’s price collapse, NAND contract prices rebounded 30–50% in 2024. But the capacity utilization is still recovering. HBF provides a new, higher-value outlet for NAND capacity, transforming flash from a commodity storage medium into a quasi-memory tier. This is exactly the kind of value migration that macro watchers look for: an asset class redefining its role in the computing stack.
Contrarian: The Decoupling Myth
Most coverage frames HBF as a direct threat to HBM. I think that’s wrong. HBF will not replace HBM in training—the physics of NAND latency is a hard ceiling. The real risk is that HBF never achieves critical mass because the ecosystem isn’t ready. No major cloud provider has committed to integrating a new memory class. The controller, firmware, motherboard interface, and AI framework support all need to be designed from scratch. SanDisk has strong IP in NAND controllers, but building an ecosystem around a new memory standard requires years of collaborative validation. HBM took nearly a decade to reach its current dominance. HBF’s window is tight: if SK Hynix or Samsung launch a “HBM Lite” product with lower cost and higher bandwidth, HBF’s cost advantage could evaporate before it even ships.
Watch the flow, not the flood. The real story here is not about HBF beating HBM. It’s about SanDisk using the macro environment—trade restrictions, inference demand, NAND cycle—to reposition itself as an AI memory innovator. The flood of HBM investment is already peaking. The flow of NAND-based memory is just beginning. But liquidity is a liar: just because capital is available doesn’t mean the technology will deliver.
Takeaway: The Signal to Track
HBF is a pure option on the AI inference scaling thesis. The bull case: by 2027, HBF captures 10% of the inference memory market, generating $5–8 billion in revenue for SanDisk, and re-rating the company from a cyclical NAND maker to a structural AI memory player. The bear case: the ecosystem fails to coalesce, HBM-dominant players respond with a price cut, and HBF becomes a footnote in memory history.
To decide which path we’re on, watch three signals. One: Does any top cloud provider announce a design win in the next six months? Two: Does JEDEC begin a standardization process for HBF? Three: Does the US Commerce Department include HBF-like NAND memory in the next export control update? If the first two happen, the flow is real. If the third happens, the game changes entirely.
Code is law until it isn’t. And in this case, the law is the physics of memory versus the economics of scale. SanDisk is betting that for AI inference, the economics win. I’m not convinced yet, but I’m watching the flow.