A new memory technology called High Bandwidth Flash, or HBF, is at the center of a strategic crossroads in the AI industry.
At its core, HBF is a clever way to bring massive amounts of memory closer to AI processors. Imagine stacking many layers of NAND flash—the same memory in your SSD—into a single, powerful package. This aims to solve the "memory wall," a major bottleneck where AI chips are so fast they spend most of their time waiting for data to arrive. HBF is designed to be a middle ground: much larger capacity than the ultra-fast HBM memory currently used, but faster than traditional SSDs.
This new technology has created a fork in the road, with two tech giants taking different paths. First is the path championed by Google, SK hynix, and SanDisk. Google's research points to AI inference becoming severely limited by memory speed and capacity. For them, HBF is a logical next step to place huge datasets (like the 'weights' of a large language model) right next to their TPU processors. This on-package approach could dramatically speed up performance for certain AI tasks.
Second is NVIDIA's path. NVIDIA believes its existing ecosystem is already powerful enough. Their strategy centers on a technology called GPUDirect Storage, which lets GPUs pull data directly from networks of high-performance enterprise SSDs (eSSDs). From their perspective, this approach is more mature, scalable, and avoids the risks of adopting a brand-new memory type. They seem to be saying, "Our current tools, when scaled up, are good enough."
Adding a twist to this technical debate is the economy. The price of NAND flash is skyrocketing due to immense demand from the AI boom. This makes building a new, NAND-heavy product like HBF more expensive and riskier. For now, it makes NVIDIA’s strategy of leveraging existing, qualified eSSDs appear more pragmatic and cost-effective.
So, the recent rumors reflect a real strategic divergence. Google is betting on a new hardware architecture with HBF to solve a future bottleneck, while NVIDIA is focused on optimizing its proven, software-driven storage architecture. While HBF is moving from concept to reality with standardization efforts, its path to mass adoption in 2026 is clouded by these technical and economic pressures. The industry is essentially watching two different philosophies compete to feed the insatiable data appetite of AI.
- HBF (High Bandwidth Flash): A new memory technology that stacks NAND flash chips to create a large, high-speed memory layer designed to be placed directly on the same package as an AI processor.
- NAND Flash: The type of memory used in Solid-State Drives (SSDs). It's denser and cheaper than DRAM (like HBM) but generally slower, making it ideal for high-capacity storage.
- AI Inference: The process of using a trained AI model to make predictions on new, live data. This is the operational phase of AI, distinct from the initial training phase.