Memory is the bottleneck for everything—AI training, data centers, your laptop. A technology that genuinely solves this would be huge. The question is whether this is a lab curiosity or something that can actually be manufactured at scale.
Researchers have developed a 3D memory architecture that combines NAND storage and DRAM performance, potentially breaking through the "AI memory wall" with faster, cheaper memory that reportedly has unlimited endurance. If it works at scale, this could reshape the entire memory industry.
That's a big "if."
According to TechRadar, the breakthrough involves reviving old camera technology—specifically charge-coupled devices (CCDs)—and adapting them for memory storage. The details are sparse, but the concept is intriguing: NAND is cheap and dense but slow. DRAM is fast but expensive and volatile. A hybrid that combines the best of both would be revolutionary.
The problem is that we've heard this before. The last dozen "revolutionary memory technologies" didn't make it to production. Memristors were going to change everything. Phase-change memory was the future. 3D XPoint (Intel's Optane) actually shipped, and it was genuinely impressive—but Intel discontinued it in 2022 because manufacturing costs were too high.
Making a technology work in a lab is one thing. Manufacturing it at scale, economically, with acceptable yields, is entirely different. That's where most of these breakthroughs die.
Still, the AI memory wall is a real problem. Training large language models requires massive amounts of high-bandwidth memory. Nvidia's H100 GPUs use HBM3 (high-bandwidth memory), which is incredibly fast—and incredibly expensive. costs more per gigabyte than any other memory type, and supply is constrained. If you could replace it with something cheaper and faster, AI training costs would plummet.
