The 6-Inch Wafer That Every AI Data Center Depends On

NVIDIA just put two billion dollars into a chip factory. It runs six-inch wafers. Not twelve. Six. So why would the biggest company in AI build on a size the industry left behind in the nineties? Because the chips moving AI's data between processors aren't silicon. They're indium phosphide — the lasers every AI data centre runs on. Coherent is spending six hundred and fifty million to quadruple that Texas line. In a silicon fab you lose yield across a thousand steps and trace it backwards. Where does compound yield go? Into the crystal itself. The defects arrive in the substrate — the money is decided at crystal growth and epitaxy, the layers grown on top, before the fab really starts. It's a furniture maker who joins everything perfectly and still loses the piece, because the flaw was in the timber before it reached the workshop. Every instinct silicon taught you points the wrong way. The bottleneck in AI was never the GPU you keep reading about. How much of it is a six-inch wafer? Now, your FabSpeak Tip of the Week. Microsoft has just previewed something in Copilot Studio that this industry solved decades ago. You can now share an agent run-only — your colleagues can run it, but they cannot change it. That is recipe management. On the floor the process engineer owns the recipe, the operator runs it, and nobody rewrites it at three in the morning. So build the boring one first — the end-of-shift summary, the excursion write-up, the thing you retype every week. Build it once, share it run-only, and your whole team inherits your version. One person's good habit quietly becomes everybody's default. That's FabSpeak — see you next week, one layer down.