Google backs SK Hynix’s HBF: “AI needs larger memory capacity”
Google backs SK Hynix’s HBF: “AI needs larger memory capacity”
FMS 2026: A solution to the memory bottleneck in the AI inference era
“The smarter AI becomes, the more data it needs to remember.”
Google DeepMind has identified High Bandwidth Flash (HBF) as a key memory technology for the next generation of artificial intelligence.
As AI evolves beyond simply generating answers and moves toward interacting with users, maintaining context, and performing complex reasoning, traditional DRAM alone may no longer be sufficient due to limitations in capacity and cost.
HBF as a new memory layer for AI
At FMS (Future of Memory and Storage) 2026 in Santa Clara, California, a panel titled:
“Breaking the Memory Wall with High Bandwidth Flash”
brought together:
Xiaoyu Ma, researcher at Google DeepMind;
SK Hynix Vice President Im Eui-cheol;
SanDisk Vice President Rajiv Nagarivara.
The discussion focused on how memory architecture must evolve for future AI systems.
Google: AI inference semiconductor market could grow more than 10x
Google DeepMind expects the semiconductor market for AI inference to grow more than tenfold over the next decade.
The reason is that the AI industry is shifting from:
training AI models,
to
inference, where AI answers user queries, performs tasks, and interacts in real time.
This shift is creating massive demand for memory.
Key drivers include:
the rise of multimodal AI (text, images, video, voice);
the growing need to store conversation history;
the rapid expansion of KV cache data used by AI models.
AI therefore requires not only faster memory but also much larger memory capacity.
What is HBF?
High Bandwidth Flash (HBF) is positioned between:
HBM (High Bandwidth Memory), the ultra-fast memory used with AI GPUs;
SSDs, which provide high storage capacity.
HBF aims to:
deliver data quickly for AI workloads;
store significantly more data than HBM;
reduce overall AI infrastructure costs.
HBF is not designed to replace HBM. Instead, it acts as a complementary memory layer that efficiently manages frequently accessed AI data.
SK Hynix and SanDisk push for a new standard
SK Hynix and SanDisk announced the first HBF standard specification at FMS 2026.
This came about six months after the launch of the HBF consortium.
Google and AI semiconductor company Tenstorrent are also participating in the standardization effort.
SanDisk said HBF could:
improve AI response times by managing conversation history more efficiently;
reduce GPU workload;
retain data even when power is turned off because it is based on NAND Flash technology.
SK Hynix: Combining HBM and HBF for AI efficiency
SK Hynix explained that future AI systems will require joint optimization of:
memory;
software;
system architecture.
The expected structure:
HBM → handles ultra-fast computation workloads;
HBF → handles large-scale data storage and retrieval.
This combination could significantly improve AI infrastructure efficiency.
Why this matters
Industry observers believe the importance of this announcement is not only the HBF technology itself.
The key point is that Google, a major AI service developer, publicly acknowledged the need for a new memory architecture proposed by SK Hynix and SanDisk.
This could accelerate the development of a next-generation AI memory ecosystem.
FMS 2026, held from August 4–6 in California, is one of the world’s largest events for memory and storage technologies.
Major memory companies, including Samsung Electronics, SK Hynix, Micron Technology, and SanDisk, participated.