The AI world is moving past the “bigger model” phase. Now it is about running those models for millions of users at once. Inference is where the money is, and inference is brutal on infrastructure. Especially memory.
That is the backdrop for SK hynix and Sandisk announcing a joint push to standardize a new memory concept called High Bandwidth Flash, or HBF. The companies held a kickoff event at Sandisk’s headquarters in Milpitas, California, and confirmed they are forming a dedicated workstream under the Open Compute Project to drive global standardization.
Going through OCP is not random. It signals that this is meant to be ecosystem-wide, not just another vendor-specific layer that locks customers in.
HBF is positioned as a new tier between High Bandwidth Memory and SSDs. HBM delivers extreme bandwidth but is limited in capacity and cost efficiency. SSDs provide large capacity but cannot keep up with HBM’s speed. HBF is designed to sit in between, filling what the companies see as a growing architectural gap in AI systems.
As AI shifts from model training to inference, the number of concurrent users spikes. That means more data movement, more memory pressure, and more demand for efficiency. According to SK hynix, existing memory structures “cannot meet the high capacity data processing and power efficiency at the same time in the inference stage,” which is exactly the problem HBF is meant to solve.
In this proposed architecture, HBM continues to handle the highest bandwidth tasks. HBF acts as a supporting layer, expanding usable capacity while helping manage power consumption. The pitch is scalability without blowing up energy budgets.
The companies also argue that HBF can reduce total cost of ownership while improving system scalability. That matters in an era where AI infrastructure spending is climbing fast, and every watt in the data center is scrutinized.
Industry forecasts suggest demand for more complex memory solutions, including concepts like HBF, could accelerate around 2030. SK hynix and Sandisk are clearly trying to define the standard before that wave hits full force.
There is also a strategic angle. In the AI inference market, competitiveness increasingly comes from system-level optimization across CPU, GPU, and memory rather than the performance of a single chip. A vendor that can deliver both HBM and HBF can present itself as a total memory solution partner.
Ahn Hyun, President and Chief Development Officer at SK hynix, put it this way: “The key to AI infrastructure is to go beyond the performance competition of individual technologies and to optimize the entire ecosystem.” He added, “Through HBF technology standardization the company will establish a cooperative system and present an AI-era optimized memory architecture to create new value for customers and partners.”
That language is about cooperation, but it is also about influence. Whoever shapes the standard shapes the market.
Whether HBF becomes a core building block in future AI servers remains to be seen. Adoption, real-world performance, and hyperscaler buy-in will ultimately decide that.
Still, as inference becomes the dominant AI workload, memory architecture is turning into a quiet battleground. And SK hynix and Sandisk are making it clear they want to define the next layer before someone else does.
