Data Centers Hit Limits on Power and Labor: AI Is Growing Faster Than Construction

Executives at some of the largest companies in the data center industry say that construction of AI capacity is barely keeping up with demand. In their assessment, the two main constraints are access to power and a shortage of skilled workers. Bloomberg Tech published the story on 13 October 2026. The available excerpt contains no specific megawatt figures, timelines or dollar amounts, so what follows examines the logic of the statement rather than the numbers of individual projects.

Where Construction Gets Stuck

"Power" here covers several different things: generation, grid connection and substation equipment. An AI cluster campus draws power comparable to a small city, and any one of these components can become the slowest link. Money and land are easier to find: a site can be bought in a matter of months, while grid connections and transformer deliveries stretch over years. That is why operators increasingly describe available power as their main asset.

The second constraint, people, gets discussed less often, even though it is felt immediately on a construction site. A data center needs electricians, cooling system installers, commissioning engineers and operations specialists. These roles cannot be filled in a quarter, and many projects are being built at the same time, all competing for the same workers.

What the Statement Says

• Source: Bloomberg Tech, 13 October 2026.
• Who is speaking: executives at some of the largest companies in the sector.
• State of the industry: construction is "barely keeping up" with the growth of AI.
• Constraint No. 1: access to power.
• Constraint No. 2: skilled labor.
• What the available text does not contain: stated capacity volumes, commissioning timelines or investment amounts.

Why This Changes the Market

A couple of years ago, accelerators were seen as the main scarcity. Now, judging by what operators say, a chip without delivered power becomes an expensive piece of cargo sitting in a warehouse. That shifts the advantage. The winners are companies that secured connections in advance, negotiated for generation and assembled contractor crews, not those that simply placed a large equipment order.

There is a second effect: when construction runs at its limit, any disruption, such as a delayed transformer or a contractor missing its schedule, immediately pushes back a cluster launch. Schedules have almost no safety margin. For model developers, this means that training and inference plans depend more and more on physical infrastructure, and not only on algorithms and budget.

What Teams and Investors Should Do

Those buying AI capacity should ask providers not only for the price per GPU-hour, but also for the site's grid connection status, the delivery schedule for electrical equipment and whether the facility has enough staff. These are more reliable indicators of real timelines than promises about the number of racks.

Teams building products on other companies' clouds should build in schedule buffers and design their systems to tolerate capacity shortages: inference optimization, quantization and flexible workload distribution across regions reduce dependence on a single campus. Investors may want to look at companies along the chain of power and skilled labor, such as training for electricians and engineers and grid equipment, and not only at chip makers. The true scale of the shortage will become clear when operators disclose project-level figures. For now, the direction is plain: the constraint has shifted from silicon to kilowatts and people.