Seven hundred billion dollars. That is the combined capex the four hyperscalers have committed to AI infrastructure this year. Amazon alone just dropped twelve billion on a single facility in Louisiana. SoftBank is planning a campus in Ohio so large its CEO casually mentioned five hundred billion dollars as if he were ordering lunch.
The bottleneck conversation always goes the same way. First it was chips. Then it was power. Then it was cooling. Then it was capital. Each time, the assumption was the same: solve this constraint and the buildout accelerates. Each time, someone found a way to throw money at the problem until it moved.
There is one bottleneck you cannot solve with money. Or rather, you can, but the money takes twenty years to produce results.
The Hands Problem
AI cannot build its own data centers. That line comes from the CEO of the world's largest recruitment firm, and it is the most important sentence in the entire AI infrastructure conversation right now.
Demand for robotic technicians is up 107% since 2022. HVAC engineers, 67%. Industrial automation techs, 51%. Plain old electricians, 27%. These are not job postings at startups hoping to get funded. These are openings at facilities that are already under construction, with concrete poured and steel going up, waiting for the people who know how to make the building actually work.
A data center is not a warehouse with servers in it. It is a precision-engineered thermal and electrical system. The cooling alone requires specialists who understand fluid dynamics, heat exchange, and increasingly, liquid cooling architectures that most HVAC technicians have never touched. The electrical systems run at densities that would have been considered absurd five years ago. The plumbing -- yes, plumbing -- is critical infrastructure when your cooling strategy involves pumping fluid through server racks.
You cannot download these skills. You cannot prompt-engineer them into existence. You cannot deploy them from a container image.
The Twenty-Year Lag
Here is the math that should terrify anyone planning a datacenter buildout timeline. A competent electrician takes four to five years to train through an apprenticeship. A specialist in high-density datacenter electrical systems takes another two to three years on top of that. A liquid cooling engineer who can design, install, and troubleshoot a direct-to-chip cooling loop for a rack pulling 120 kilowatts does not come out of a trade school. They come out of experience.
The pipeline for these people was not built for this moment. The pipeline was built for a world where datacenters were a niche construction category, not the single largest driver of commercial construction spending in the developed world.
So what happens? Wages go up. HVAC engineers are seeing 10 to 15 percent increases. Specialists moving into datacenter roles are getting 25 to 30 percent bumps. Six-figure salaries for trade workers are becoming normal. This is the market working exactly as it should -- scarcity creates a premium, the premium attracts talent, and eventually supply catches demand.
Eventually. That word is doing a lot of heavy lifting.
The Uncomfortable Truth
We have spent the last three years talking about AI replacing jobs. White-collar anxiety is at a fever pitch. Meanwhile, the entire AI revolution is physically dependent on people who work with their hands, and there are not enough of them.
The irony is sharp enough to cut. The technology that threatens to automate knowledge work cannot exist without the one category of labor it cannot automate: the physical construction and maintenance of the infrastructure it runs on. Every neural network, every large language model, every inference call lives inside a building that someone had to wire, plumb, and cool by hand.
A fourth-generation plumber described the modern datacenter as the first workplace where highly compensated trade workers physically work alongside network engineers with college degrees, valued the same. He called them new-collar jobs. I think he is underselling it. These are not new-collar jobs. These are the jobs the entire digital economy literally cannot function without.
What This Actually Means
If you are building datacenters right now, your timeline is not constrained by how fast you can get GPUs from NVIDIA or power from the utility. It is constrained by how many qualified people exist who can turn a construction site into a functioning facility. And that number is not growing fast enough.
If you are a twenty-year-old trying to figure out what to do with your life, I would point you at a five-year apprenticeship in electrical or mechanical trades with a datacenter specialization before I would point you at a computer science degree. The CS graduate is competing with every other CS graduate on the planet plus an increasingly capable set of AI tools. The datacenter electrician is competing with a talent pool so small that companies are offering signing bonuses that used to be reserved for software engineers.
You can always build more chips. You can always generate more power. You can always raise more capital. You cannot fast-forward a human being through ten years of hands-on experience.
That is the last bottleneck. And it does not have a software fix.