AI data centers are about to spend $1.6 trillion. Building them is the hard part.

AI data centers are projected to absorb $1.6 trillion in capex by 2031, but the constraint that decides whether that money becomes operational capacity is not GPUs or gigawatts. It is whether precision manufacturing, the glass interposers, co-packaged optics, and advanced PCBs that move data between chips, can scale fast enough to keep up with the build-out clock.

The freshest signal sits inside the optics market. TrendForce forecasts co-packaged and near-packaged optics connections will grow from roughly $100 million in 2025 to more than $39 billion by 2030, a near-four-hundred-fold jump driven by AI's appetite for bandwidth between accelerators (TrendForce, June 2025). That is the size of the category AI has to build.

The Goldman Sachs capex ceiling makes the tension explicit. The bank projects AI infrastructure spending will rise from $765 billion in 2026 to $1.6 trillion by 2031 (Goldman Sachs). The same report flags power, labor, and equipment as the bottlenecks that will lengthen the gap between committed capital and switched-on data centers. Equipment is the variable that gets the least public attention, and the precision-manufacturing layer is where the equipment constraint lives.

A chiplet package today is a stack of small silicon dies that have to talk to each other at terabit speeds. They do that over microscopic copper-filled channels drilled through a thin sheet of glass, a glass interposer, and through optical interconnects that replace electrical traces with light, letting data move between chips at far higher bandwidth and lower power. Drilling those through-glass vias and trimming the surrounding glass and PCB material is delicate work. A femtosecond laser fires pulses a quadrillionth of a second long, so short that the material being cut barely heats up. That lets manufacturers drill and scribe glass and other substrates without cracking them, which older thermal lasers cannot do at production volumes.

The demand for those tools is now outrunning supply. A LITILIT representative told Business Review that order books have lengthened and capacity has become the binding question for several of the company's customers (Business Review). Independent trade press has confirmed the company's response: a new Vilnius factory is being brought online to scale production (Electronics Weekly, BIS Infotech, Eureka Magazine). LITILIT is one vendor among several, and the claim that femtosecond-class tools are a market-wide bottleneck is its own framing, not an analyst consensus. The mechanism it points to, a narrow set of precision tools sitting between AI demand and shipped hardware, is structural.

The macro variable is the TrendForce adoption curve. Co-packaged and near-packaged optics are not yet the dominant interconnect architecture in shipped systems. If CPO and NPO ramp faster than TrendForce expects, the precision-manufacturing layer has to scale ahead of the curve, and laser capacity, glass-via drilling throughput, and optical assembly lines become the rate limit. If they ramp slower, the constraint moves back to packaging yield and substrate supply, and the laser story fades.

The clock matters because capex commitments are public while the manufacturing ramp is not. Hyperscaler procurement teams have already moved to multi-quarter order books for advanced packaging tools, and substrate suppliers are quoting 2027 capacity. The variable to watch over the next two quarters is whether femtosecond-laser and glass-via drilling throughput grow on the same curve TrendForce's CPO and NPO forecast implies, or fall behind it.

Goldman's $1.6 trillion is a capital number. Whether it turns into running data centers is a manufacturing question, and manufacturing is where the next eighteen months of AI infrastructure will be decided.