AI Infrastructure Runs on Light, Not Just More Chips · Hanh D. Brown

The model you talk to is the tip of the iceberg. Under the water sit energy, chips, and miles of optics that carry data as light. Most coverage of Artificial Intelligence (AI) stops at the part you can see. The build-out underneath is the real story.

Why is AI called a new industrial revolution?#

A chatbot is not the machine. It is the dashboard on the front of one. Behind it runs an enormous physical plant, humming in a building you will never visit. That plant is where the money and the risk actually live.

The stack has five layers.

• Energy at the bottom, powering it all.
• Chips above, the raw compute.
• Data centers, the land and cooling.
• Cloud and models, the software.
• The apps you touch, on top.

Every answer you get has climbed up through all five.

Like an iceberg, the chatbot is the tenth above the waterline. The rest is a supply chain of power plants, fabrication lines, and cooling systems. Real steel. Real power. All of it has to be built and paid for. Intelligence is a general-purpose engine now. It reaches into healthcare, transport, and the robots on the factory floor.

That is why the industrial-revolution comparison keeps coming up. It is not one product. It is a base layer, and it reshapes what sits on top the way cheap steel and cheap power once did. This is the same shift the reinvention of computing has been forcing for a decade. The wonder is real. It is also only the visible tenth.

Why do AI data centers need optical interconnect?#

Optical interconnect is the part most coverage skips. Here is the line to keep: AI runs on compute, but it scales with connectivity. A large system is never one chip. It is hundreds of thousands of them, all forced to behave as a single machine.

Over a short run inside a board, copper wire carries the signal fine. Push it out to hundreds of feet, across a field of racks, and copper fails. At that range, light is the only practical way to move the data.

Mechanically, it is clean. A small current fires a laser, and the light runs down a glass strand thinner than a hair. At the far end a detector turns it back into current, which drives the next chip. Current, to light, to current again.

So the parts that carry light get treated as core infrastructure rather than spare cable in a closet. When your whole machine is a million processors pretending to be one, the thing that lets them talk is the constraint. Compute gets the headlines, but connectivity decides how far the compute can actually reach.

What is an AI factory?#

An AI factory is a plain thing dressed in a big name. It is a very large cluster of computers. It is built to run one huge job as if it were a single machine. The work gets split and spread across the whole floor.

That splitting has a name. It is called sharding. Thousands of processors take a slice each and run in parallel. At full size it is the largest, most compute-hungry program ever run. Like a foundry pouring one white-hot batch, the plant bends toward a single output.

Picture what that buys in research. You ask a question. The system reads a stack of documents. It chases the references. It grounds itself on what it can verify before it reasons. Then it can direct lab robots to run an experiment and hand back a measurement.

That loop, read then verify then act, is offered as a possible future for drug discovery. Keep the word factory in perspective, though. It is a vivid handle, not a fixed standard. Treat it as a picture, not a spec.

What is indium phosphide used for in AI data centers?#

Indium phosphide is the material that makes the light. The tiny lasers inside an AI system are built from it. Those lasers are what send data between distant chips.

Run a small current through an indium-phosphide laser. It emits light. That light carries information across gaps that copper cannot bridge. So this one metal sits on the critical path for wiring a large system together. That is why a fab that makes it draws so much attention.

This facility makes six-inch wafers of the stuff. Its operators call it the first and largest of its kind, which is worth marking as a company claim rather than an audited fact. The plan announced is to double the site and quadruple its output.

One number travels with the pitch. It is the demand signal to watch. It took roughly fifty years to build one unit of this capacity. The plan is to build four times that in a single year. Read it as the company’s figure, not a law. The direction is the point.

Can AI bring back American manufacturing jobs?#

Manufacturing jobs are the stated hope here. The numbers behind that hope come from interested parties. So hold them loosely. The honest version keeps the mechanism and marks the figures as projections.

Its argument runs simply enough. AI demand is pulling new chip fabs, packaging plants, and data centers onto American soil. That build-out needs hands. Skilled ones. The figures floated include roughly 600,000 jobs from recent construction. At the one site in question, the claim is more than a thousand direct and indirect roles.

Those are forecasts and company statements, not audited results. Treat them as direction, not fact. What is sturdier is the strategy underneath, and it is where the real economic value sits.

You rarely revive an old industry by fighting it head-on. Like a new road cut through open country, a genuinely new industry makes room where an entrenched one had none. The bet is plain. Advanced manufacturing returns not by reopening the old plant, but by riding a new one. Plausible. Still unproven at scale.

How does AI demand affect energy and the power grid?#

Start at the bottom of the stack. That is power. A data center is a very large electricity customer. A field of them is a planning problem for a whole region.

Optimists make a market case. Steady demand from AI gives a reason to build more generation, including solar and nuclear, funded by private money rather than public subsidy. That is the builders’ argument, and it carries their commercial interest. So read the no-subsidy part as advocacy, not settled fact.

What is concrete is the load. These machines draw serious power. Meeting that draw is now a central question for the industry and the utilities behind it. Energy is the precondition. Without the power, the fabs and the data centers do not run at all.

So bring it to the kitchen table, because this is where it touches a family. The same grid that feeds a data center feeds your home, and the bill that lands each month rides on how well a region plans its power.

If your work or your household leans on these systems, the supply of energy and optical parts is a real risk, not a distant one. Watch where the new power gets built this year. That is the story worth following.

Source: Jensen Huang of Nvidia and the chief executive of Coherent, in a facility conversation in Sherman, Texas.