If you’ve been following AI news lately, you’ve probably heard Jensen Huang mention a new bottleneck: inference. In plain English, while we used to focus on “training” AI to learn, we are now focused on how AI “thinks” and responds to us in real-time.
To do this, AI needs a massive amount of memory.
The “Software Fix” Fallacy
Recently, Google released a program called TurboQuant that makes AI software use 6x less memory. You might think this would mean companies will buy fewer memory chips. Actually, the opposite is true.
Think of it like a closet. If you suddenly find a way to fold your clothes 6x smaller, you don’t move into a smaller house. You just buy 6x more clothes.
In the tech world, when we make something more efficient, we don’t save—we explore possibilities to expand. Tech companies will always maximize their hardware to “future-proof” their chips, meaning the demand for memory isn’t going anywhere.
Meet HBF: The AI Marathon Runner
Until now, the star of the show was HBM (High Bandwidth Memory)—it’s incredibly fast, like a sprinter, but it’s small and expensive.
The next big thing is HBF (High Bandwidth Flash). If HBM is a sprinter, HBF is a marathon runner. It’s a new type of memory that is:
- Massive: It can hold 10x more data than HBM.
- Affordable: It costs much less to produce.
- Efficient: It uses 40% less power.
As AI models get bigger, they need a “giant library” that they can search through quickly. This is where companies like SanDisk come in. By leading the charge on HBF, they are turning “storage” into the engine that will power the next generation of AI.
The bottom line? Better software won’t kill the memory market; it’s just clearing the way for a bigger, faster future.
Internet of Things and More 
Leave a comment