A data center infrastructure insider tells For Humanity that supply chain limits, gas turbine backlogs, and a skilled trades shortage could slow the AI buildout dramatically. Here is what that means for AI safety.
Most of the debate about how fast artificial intelligence will advance happens in software. Benchmark scores, model releases, and the shape of the capability curve dominate the conversation. But there is a second clock running underneath all of it, and almost no one on the capabilities side is reading it: the physical timeline for building the AI data centers that make any of this possible.
In a recent episode of For Humanity, host John Sherman spoke with Jon Billow, who is on the leadership team at BNS, a firm that manufactures and installs the electrical and critical power infrastructure behind large data centers. According to Billow, the AI buildout cannot move at the speed the major labs imply, and the gap between the promised timeline and physical reality is large, measurable, and mostly ignored.
His estimate is blunt: whatever AI infrastructure timeline you have been handed, you may need to multiply it by five to seven.
Billow frames the problem using the theory of constraints. Any system moves only as fast as its weakest link, and a large data center is a system with many links that all have to arrive at the same time. He lists permitting, grid interconnection, critical power, cooling, and compute. Miss one, and the project waits.
The link he returns to most often is critical power equipment, because that is where the data center supply chain narrows almost to nothing. Billow says the orders effectively funnel back to roughly five manufacturers, naming Eaton, ABB, Schneider, and GE Vernova among them, and that all of them carry backlogs measured in years. He notes that even the US government struggles to get its allocation for Navy and Coast Guard ship programs, because it is now standing in the same queue as every hyperscaler racing to expand AI capacity.
The pressure does not ease from there. Billow points out that more municipalities now require data centers to bring their own behind-the-meter power generation rather than simply drawing from the grid. He supports the logic, but argues it stacks a second equipment backlog on top of the first and forces operators to develop a skill they have never needed: assembling gas turbines and finding electricians who can parallel generators. The skilled trades, he says, are among the most constrained labor markets in the entire picture.
When Sherman pressed him to put a number on the gap, Billow offered a rough calculation. By his account, the US has roughly 50 gigawatts of total data center capacity, with about a quarter of it allocated to AI. Around five gigawatts are under active construction, with another seven to twelve gigawatts in backlog. Set that against the order-of-magnitude jumps in capacity the labs describe as necessary for the next tier of capability, and his read is that real timelines may run five to seven years where companies are promising one.
It is worth noting that Billow's claim does not stand alone. Industry reporting on the gas turbine backlog has independently described delivery waits stretching from roughly two to three years a few years ago to as long as five to seven years today, driven heavily by data center demand. An insider's anecdote and the wider market data point in the same direction.
Billow highlighted two recent developments that, in his reading, suggest infrastructure is already the limiting factor. He pointed to OpenAI reportedly pulling back from a large commitment tied to its Sora video product, which he interprets as a company looking at finite compute and deciding where to spend it. He also referenced a delayed model release from Anthropic, which he attributes partly to legitimate security concerns and partly to constrained compute capacity.
His broader point is that software capability keeps leapfrogging forward, but the ground underneath it does not move at the same speed. He also closes one common escape hatch. When asked whether more efficient compute might reduce the need for buildout, he invoked Jevons paradox, arguing that as a resource becomes more efficient, demand for it tends to rise rather than fall.
Billow does not frame any of this as a reason to relax. He frames it as time. If the physical AI infrastructure runs years behind the hype, that delay is runway to get AI governance and alignment right rather than scrambling after the fact. He draws a parallel to how the world slowly built doctrine around nuclear risk, and argues the task now is to use the delay deliberately rather than waste it.
He was candid about the limits of his optimism. On the question of whether humanity can control a system far smarter than itself, Billow agreed at a conceptual level that this is genuinely hard, while holding open the possibility that immutable constraints built into a model from the start could help. He stressed he does not know how to guarantee that, calling it a hope rather than a solution.
The conversation did not paper over tension. Sherman described his recent time in Holly Ridge, Louisiana, standing beside a data center he compared in scale to Manhattan, in a town of roughly 2,000 mostly elderly residents, with construction dust in the air and water residents will not drink. He found it overwhelmingly sad.
Billow, who works around these structures constantly, sees them differently. He describes a data center as a testament to human progress and ingenuity, no more inherently harmful than an automotive factory, and argues the real issues are siting and governance rather than the technology itself. He also made a useful technical point: because AI training and much inference are not geographically bound the way older workloads were, there is little reason these facilities cannot be built in remote locations like the desert rather than next to communities.
Both reactions are real, and the episode is stronger for holding them in the same room.
Congressman-style policy fixes and lab-level alignment research both matter, but Billow's contribution is a reminder that the AI buildout is also a physical, industrial process with real limits. As he put it, he wants to tell his grandkids that we were building the car while it went down the road at 55 miles an hour, but we had the presence of mind to put in seatbelts because we knew who was in the back seat.
The seatbelts do not install themselves. If the supply chain is buying us time, the work now is to spend it on the safeguards, governance, and public awareness that AI safety depends on.
Watch the full conversation with Jon Billow on The AI Risk Network: https://www.youtube.com/@theairisknetwork
Take action on AI risk: https://safe.ai/act
