Is AI Becoming Part of Its Own Supply Chain?

AI is already helping to build the next generation of AI models. But it’s also moving further down the stack, into the machines building the chips they run on. In May, NVIDIA announced that TSMC is using vision AI to improve defect inspection inside its semiconductor fabs. And Samsung is building entire digital twins of its semiconductor factories to identify anomalies and predict maintenance problems before changes are made in the real world. So, AI isn’t just helping to build better AI. It’s starting to monitor the physical machines that make it possible. AI Is Already Inside the Chip Factory Semiconductor manufacturing produces a ridiculous amount of data. NVIDIA says TSMC is dealing with hundreds of thousands of process parameters across thousands of manufacturing steps. So, it’s not surprising that chipmakers are turning to AI. TSMC is using NVIDIA’s Metropolis platform and TAO Toolkit for automated defect inspection. NVIDIA says the system can improve the detection of nanometer-scale defects. KLA makes inspection equipment used by semiconductor manufacturers. Its systems use deep learning and machine learning to separate tiny defects from the patterns and background noise around them. AI is getting better at finding things humans might miss. But finding a defect is only half the problem. Which Defects Actually Matter? Let’s say an inspection system flags two defects. One is clearly bigger than the other. So which one is more dangerous? You’d probably pick the bigger one. But that isn’t always how it works. A recent experiment by optical manufacturer OPTOMAN and inspection specialist DIOPTIC looked at microscopic defects in high-power laser optics. The researchers compared visible surface defects with areas of high absorption and the locations where laser damage eventually occurred. The results were surprising. Three damage sites were identified. One was associated with a relatively large defect of around 5 microns. But the other two were linked to smaller defects. Those smaller defects also happened to sit at the two highest absorption points found during testing. So the biggest visible defect wasn’t necessarily the most dangerous. Surface defects, absorption hotspots and laser-damage sites from the OPTOMAN and DIOPTIC test. That creates a problem for automated inspection. And that problem isn't unique to laser optics. The more AI gets used for industrial inspection, the more useful it becomes to distinguish between something that looks unusual and something that is likely to fail. Another test from the same research shows how important these tiny imperfections can be. A single high-absorption defect reduced the measured laser-induced damage threshold by more than 40%. That means the optic was able to withstand significantly less laser energy because of one localized problem. That’s a huge impact from something you might barely be able to see. And it highlights the wider problem for AI-powered inspection. Finding a defect doesn’t necessarily tell you how much it matters. Can AI Spot Problems Before They Happen? Samsung is working with NVIDIA to build what it calls an AI Factory. Using NVIDIA Omniverse, Samsung is creating virtual versions of its fabs. These digital twins can identify anomalies, perform predictive maintenance and test changes before they are introduced into the physical factory. The aim is to move from asking what went wrong to asking what might go wrong next. And that could eventually apply to much smaller components too. Instead of flagging every microscopic defect, AI systems could help engineers work out which ones deserve the most attention. The Hard Part Is Getting Enough Failures If you’ve spent any time with ChatGPT or Claude, you know AI gets better when it has more data to learn from. That creates a problem when you’re trying to predict equipment failures. AI needs examples of things actually failing. Manufacturers generally try pretty hard to stop that from happening. And when failures are rare, expensive or difficult to reproduce, useful training data becomes much harder to collect. In the OPTOMAN experiment, researchers had to combine automated inspection, absorption measurements and physical laser testing to work out which defects were actually linked to damage. There’s also no guarantee that the same defect behaves the same way under different conditions. Change the material, temperature, power or manufacturing process and you may get a different result. So, even if AI spots a pattern, physical testing still matters. AI Is Becoming Part of the Process That Builds More AI AI is slowly becoming involved in more of its own supply chain. It helps write code. It helps design chips. Now it’s moving into the factories that manufacture those chips. Predicting which parts of those factories are likely to fail could be another step. That doesn’t mean AI is literally building itself. But the feedback loop is getting harder to ignore. The technology used to build the next generation of AI is increasingly being improved by AI itself.