AI Has Brain. Giving it Body is Harder Than it Sounds: Why AI Infrastructure is Running into Manufacturing Wall

OpenAI, Meta, Anthropic, and other tech companies are pouring billions into AI models that can write code, reason through problems, and hold a conversation, giving AI a sharper brain every month. But a brain needs a body, which consists of data centers, chips, and other AI infrastructure. And building one is proving challenging, partly because the AI body requires a level of precision that conventional tools struggle to deliver.

This year, AI providers are expected to spend nearly $1 trillion on AI infrastructure and a further $6 trillion between 2027 and 2030, according to a recent report from Gartner. Much of the money will go into the hardware inside data centers: AI chips and the high bandwidth memory stacked beside them, the glass substrates they are packaged on, the optical fiber and photonic components that link thousands of chips together, as well as the cooling that keeps them running.

Manufacturing the hardware is technically challenging and can lead to bottlenecks that last years. Take through-glass vias (TGVs). TGVs are tiny holes drilled through glass and filled with metal so electrical signals and power can pass vertically through it. Chipmakers are relying on them more and more, and they are notoriously hard to make well. Drilling can leave microscopic cracks and holes of uneven size, and TrendForce does not expect glass core substrates to reach commercial-scale production until after 2030.

This is where femtosecond lasers come in. They fire pulses lasting a quadrillionth of a second, so short that material is removed before heat can spread into the glass around it. The result is a clean hole with no cracks or melting, repeated thousands of times across a single panel.

Glass is only one example. High bandwidth memory is built by stacking 12 to 16 chips, each thinned to a few tens of micrometers, thinner than a human hair and easily cracked by a mechanical saw. Meanwhile, AI circuit boards are made of hard, low-loss materials that wear out drill bits quickly.

Co-packaged optics need microscopic light-guiding channels inside glass to bring optical connections right up to the processor, and as chips get hotter, engineers are building cooling channels into glass and silicon so liquid can flow close to the hottest spots.

In each case, femtosecond lasers can do work that conventional tools struggle with: cutting wafers thinner than a hair or drilling holes in hard board materials with extreme precision.

AI’s body also extends into the devices in our pockets: smartphones, laptops, and smart glasses increasingly run AI models on device. Femtosecond lasers help make them, cutting cover glass, camera lenses and display parts without cracks. Beyond AI, these new generation lasers reshape the cornea in eye surgery and serve a growing list of other fields. Yet most femtosecond lasers grew out of scientific instruments, and they remain difficult to produce in large numbers.

At LITILIT, we set out to close that gap. Our laser architecture is protected by two patents, one developed together with Lithuania’s Center for Physical Sciences and Technology (FTMC). It was designed for industrial production from the start: modular design, reduced component complexity, and a high level of automation. This is what will allow us to produce 3,000 lasers a year in our new factory in Vilnius at full capacity, the model we are planning to replicate in other countries as well.

The brain of AI was built in software labs. Its body is being built on factory floors, one microscopic hole and one crack-free cut at a time, and how fast AI grows from here depends on how quickly precision manufacturing can scale.

****