We didn't expect the biggest debate at SanDisk's Investor Day to be about a benchmark. But here we are, staring at a slide that compares High Bandwidth Flash (HBF) to High Bandwidth Memory (HBM), and it feels less like a technical breakdown and more like a political campaign ad.
Decentralization is not a tech stack; it's a philosophy of transparency. And right now, SanDisk's HBF benchmark is anything but transparent. The numbers are cherry-picked, the frame is a trap, and the bull market is the perfect fog for this kind of narrative warfare.
Context: The Battle of the Memory Layers
Let’s strip this down. HBM is DRAM-based, JEDEC-standardized, and the undisputed king of AI training. It delivers nanosecond latency and bandwidth that scales with every generation (HBM3E → HBM4 → HBM4E). HBF, on the other hand, is SanDisk’s NAND flash-based answer. It’s not a DRAM killer—it’s a capacity arbitrage play. Think of it as a high-bandwidth storage layer, not a memory replacement.
SanDisk’s presentation compared the two using a specific GPU configuration: 8 stacks of HBM at 12.8 TB/s bandwidth and 192 GB capacity. Their conclusion? HBF could reduce the number of GPUs needed for inference on a 480B-parameter model like Qwen3-480B-A35B.
Core: The Ethical Algorithmic Framing of a Benchmark
Here’s where the math gets interesting. SanDisk chose a low-end HBM configuration—HBM3E 12-high stacks at 1.6 TB/s per stack. That’s a fair baseline, but only if you ignore the roadmap. Analyst Zephyr from Citrini called out the real issue: by 2026, HBM4E 16-high stacks will deliver 4 TB/s per stack, totaling 512 GB and 32 TB/s for an 8-stack GPU. That’s triple the bandwidth and 2.7x the capacity.
But the real twist is quantization. SanDisk’s demo assumed bfloat16 precision, which makes a 480B-parameter model a memory hog. Switch to FP4/FP8—the industry standard for inference in 2025—and that 480B model drops to 240-480 GB. Suddenly, HBM4E’s 512 GB capacity covers the entire model, negating HBF’s core advantage.
SanDisk’s “12.8 TB/s” figure is a static snapshot. It’s like comparing a 2023 car to a 2026 concept car and declaring the 2023 model obsolete. The hidden agenda here is simple: SanDisk is selling a capacity story, but the market is buying a bandwidth and latency one.
Contrarian: The Pragmatic Risk of the NAND-DRAM Proxy War
Here’s the counter-intuitive take: SanDisk isn’t trying to replace HBM. They’re trying to create a new category—a high-bandwidth flash tier for inference memory pools. The real competition isn’t HBM; it’s CXL memory expansion and smart SSD offloading. The “HBF vs. HBM” framing is a marketing misdirection that masks the real technical story.
Let’s talk about the elephant in the room: the DRAM-NAND industrial divide. HBM is a DRAM oligopoly’s cash cow. SK hynix, Samsung, and Micron control the supply chain, with TSMC’s CoWoS packaging as the bottleneck. SanDisk, as a NAND IDM, has no seat at that table. HBF is their attempt to carve out a position using their existing flash fab capacity, but the high-bandwidth interface logic, TSV/hybrid bonding, and controller design demand significant new capital expenditure. The barrier isn’t NAND—it’s the packaging and the controller.
From a regulatory angle, this has geopolitical implications. HBM is caught in the US-China export control crossfire. If HBF can be manufactured with domestic equipment and bypass the HBM ban, it becomes a lifeline for Chinese AI chipmakers. But the advanced packaging bottleneck remains, and SanDisk has no incentive to solve it for geopolitical arbitrage.
Takeaway: The Vision Forward
Open source isn’t just about code; it’s about data. And in the battle for AI inference memory, the side with the most transparent benchmark wins. SanDisk’s HBF is a real innovation, but its current benchmark is a disservice to the technology. If they want to win, they need to stop comparing their flash to yesterday’s DRAM and start comparing it to tomorrow’s data pipeline.
The real question isn’t “Can HBF beat HBM?” It’s “Can the industry afford a multi-tiered memory hierarchy where NAND handles capacity and DRAM handles speed?” The answer is yes, but only if we stop pretending HBF is a substitute for HBM. It’s a complement. And the bull market, with its euphoria over AI, is the perfect time to sell that narrative—but only if the code is honest.
Art isn’t who owns it; it’s who verifies it. And in this case, the verification is still pending.