Labor Is the Bottleneck AI Can't Automate Away

Crusoe, the cloud provider staffing the Abilene, Texas datacenter, raised wages 30% to recruit enough tradespeople, and the project peaked at more than 9,000 workers on site. That single number is the structural bottleneck behind the AI buildout: there are not enough electricians and pipefitters to wire traditional buildings fast enough, so the buildings are now being wired in factories and shipped on trucks.

That constraint has turned modular construction into the default playbook across the industry, from hyperscalers to colocation providers to AI labs. According to SemiAnalysis's Modular Tracker, more than 61 gigawatts of modular capacity sit across 1,000+ sites, and the firm projects modular construction will account for 30% or more of total live datacenter capacity by the end of 2028 (The Wild Wild West Of LEGO Datacenters). "Modular" here is concrete. It means finished concrete panels, factory-wired mechanical and electrical rooms, and entire data halls rolling in on flatbed trucks. A single "brick" weighs 50,000 pounds.

Meta was the first hyperscaler to act visibly on the labor math. Over a year ago, SemiAnalysis flagged Meta's move to non-traditional membrane and fabric "tent" structures, a category the analyst firm described at the time as a drastic shift in datacenter construction. AWS has now followed with its own modular design, codenamed "SAMDC" (Scalable Amazon Modular Datacenter), which the firm reports is being rolled out at very large scale. The move from one-off experiment to default playbook at the two largest cloud companies is the inflection point. When the constraint is trades, the response is factory.

What modular does and does not solve is worth separating. On the supply side, SemiAnalysis estimates that suppliers like Vertiv can deliver roughly twice the content (the equipment, cabling, and mechanical systems inside a datacenter) per megawatt of capacity under a modular build compared to a stick-built one, because the work shifts from outdoor job sites to indoor production lines. The site also goes up faster, since the bulk of the wiring and plumbing happens off the critical path of the build. What modular doesn't solve is the upstream scarcity. Factory lines still need electricians, welders, and QC technicians. The trade shortage is being absorbed into the supply chain, not eliminated. SemiAnalysis is also skeptical that every "modular" claim in vendor marketing is real. The "Wild Wild West" label is the firm telling readers to read the next press release with a stopwatch and a bill of materials, not a brochure.

The supplier rotation is the second-order effect. If content per megawatt roughly doubles for the suppliers that win the modular build, and the buildout runs at the scale SemiAnalysis's 61-gigawatt tracker now implies, then the supplier mix that emerges from this decade will look very different from the one that built the cloud. Vertiv is the name SemiAnalysis flags today, but the firm's broader Industrials Model is built to track which other vendors absorb the same uplift. The investor question and the operations question collapse into the same one: which factory is actually shipping, and which "modular" announcement is a render.

The reader-useful version of this story is not that AI buildouts are accelerating, or that they are delayed. Both will keep happening. The useful version is that the timeline is gated by trades, and the next time a hyperscaler announces a "modular" build, the right questions are how many factory-wired megawatts they are actually shipping, how many electricians that absorbed, and which supplier captured the doubling of content per megawatt. The vendors who figure that out first are the ones who will set the next capex cycle's effective ceiling.