The hidden work keeping your AI habit running

South Korea is committing over $550 billion to build new "fabs"—the ultra-clean, high-tech factories where memory chips are made. These chips provide the short-term storage (technically called RAM—Random Access Memory) that lets an AI "think" quickly. At the same time, a company called Omen AI just raised $31 million to monitor the liquid cooling systems that keep data center computers from overheating, specifically looking for "biological" problems like bacteria in the water. Meanwhile, a ranking service called LMSYS Arena, which lets people compare AI responses to see which one works better, has turned into a $100 million business as companies scramble to prove their software is the top performer.

To understand why this matters, think of an AI model—a complex math program trained on billions of lines of text—like an incredibly well-read librarian. When you ask the librarian a question in the "cloud," they aren't actually floating in the sky. They are sitting inside a massive warehouse full of computer "servers." To give you a fast answer, the librarian needs a massive desk to spread out their "notes." This is what the memory chips in South Korea do; they provide the workspace for the AI to hold information while it processes your request. If the data center's memory "desk" is too small, the AI becomes slow and "forgets" the context of your conversation. Because these "librarians" are essentially thousands of computer processors running at the same time, they generate so much heat that they must be submerged in or surrounded by liquid cooling systems. If those systems get clogged with "biologicals"—literally bacteria or slime growing in the pipes—the whole warehouse can overheat and shut down.

We are currently in a "gold rush" phase, where the limiting factor for AI isn't just the software itself, but how many physical machines we can jam into a cooled room. By tracking these "picks and shovels" investments—the memory chips, the cooling systems, and the "leaderboards" that rank performance—we can see where the bottleneck actually lies. It turns out the biggest roadblock to smarter, faster AI isn't just better math; it's physics, pipe-fitting, and having enough physical hardware to hold all that thought.