Interests

At some point in 2023, the collective semiconductor industry woke up and realized that AI accelerators need memory bandwidth in amounts normally associated with a small country's internet backbone. The solution? Stack DRAM dies on top of each other until the thermal density approaches the surface of the sun, then call it HBM3E and ship it.

The current state of play: every HBM supplier is sold out through 2027. NVIDIA is buying every micron (pun intended) of capacity. SK hynix achieved HBM market leadership by simply not messing up their TSV process, which turns out to be harder than it sounds. Samsung is doing great if you ignore their HBM3E qualification delays. Micron is trying really hard and that's what counts.

Building a leading-edge fab costs approximately $20 - 30 billion and takes 3 - 5 years, assuming no geological surprises, trade embargoes, or unexpected shortages of ultra-pure water. By the time the fab comes online, the node it was designed for is already two generations behind, but don't worry - the industry just renames everything so your 3nm fab is actually 2nm now. Problem solved.

EUV lithography tools, each costing around $200 million and requiring a vacuum system that could double as a space simulator, have a lead time measured in quarters. ASML ships roughly 40 - 50 High-NA EUV tools per year. TSMC buys most of them. Samsung and Intel fight over the remaining ones like siblings arguing over the last piece of cake, except the cake costs a quarter billion dollars and requires its own cleanroom.