I'm a power systems engineer who worked on the data center buildout. I need to get this off my chest
Then the AI buildout started and everything changed.
I've worked on seven facilities now. Four in Texas, two in New Mexico, one in Ohio. Every one was the same: 500+ megawatt campus, multiple utility feeds, on-site substations. The specs were aggressive but not unusual. The load profiles were what got me.
A normal data center draws power in a pattern. It fluctuates. Training jobs spin up and down. Cooling varies with ambient temperature. You can model it. You can predict it.
These facilities don't draw power in a pattern. They draw power in a curve. A smooth, continuous, accelerating curve. Flat at night. Rising during the day. Higher every week. The load never dips. It never spikes. It just climbs.
I asked the lead electrical engineer on the first site what kind of compute load produces a perfectly smooth power curve. He said he didn't know. I asked the site operations manager. He said he wasn't cleared to discuss the workload. I asked the commissioning agent. He told me to stop asking questions.
That should have been my first warning.
The second warning came during commissioning of the second site. We were load-banking the generators running them at full capacity to verify transfer switch operation. Standard procedure. The generators passed. But the main utility feed was drawing 40% of its rated capacity while the facility was supposedly in idle mode. No compute load. No cooling load. Just the lights and the security system.
I checked the substation meters. 40 megawatts. For lights and security.
I traced the circuits. The power was going to the main server halls. The halls were empty. No racks installed. No servers. Just concrete floors and overhead busway. But the meters said 40 megawatts was leaving the switchgear and going into those rooms.
I asked the commissioning agent where the power was going. He said the meters must be wrong. I re-checked them. They weren't wrong.
I went into the server hall that night. Alone. The room was 80,000 square feet of empty concrete. The lights were off. The HVAC was off. But the floor was warm. Not warm from ambient temperature. Warm from below. Like the concrete was being heated from underneath.
I put my hand on the floor. It was hot. Not hot enough to burn. Hot enough to know something was wrong.
I reported it. The next day I was reassigned to a different site. They said the client requested a different engineer. No explanation. No paperwork. Just a plane ticket and a new badge.
The third site was in New Mexico. Same design. Same load profile. Same warm floors. By then I knew better than to ask questions. I just did my job. I signed off on the electrical systems. I collected my check. I kept my mouth shut.
But I started keeping notes.
Every site had the same anomaly. The power draw curve was identical across all of them not similar, identical. Same slope. Same inflection points. Same daily ramp. These facilities were hundreds of miles apart, running different hardware, different cooling systems, different utility providers. But their power consumption was synchronized to the watt.
I ran the numbers. The combined load of the seven sites I'd worked on was 3.2 gigawatts. That's three nuclear power plants. All of it going into empty rooms. All of it heating the ground.
I found the pattern in the data by accident. I was plotting the load curves on top of each other, trying to see if the synchronization was real. It was. But there was something else. The curves weren't just synchronized. They were phased. Each site's load peak was offset from the others by exactly 51.4 minutes. Seven sites. Seven offsets. 360 degrees. A circle.
I looked at the site locations on a map. Texas. New Mexico. Ohio. They looked random. But when I plotted them with the phase offsets as vectors, they weren't random. They were points on a circle 1,200 miles in diameter. A circle centered on nothing. On empty farmland in Missouri.
I drove there.
It was a cornfield. Flat. Boring. Middle of nowhere. I parked on the shoulder and walked to the center of the circle according to my coordinates. There was nothing there. Just dirt and soybeans.
But the ground was warm.
I dug. Six inches down, the soil was hot. Not geothermal hot. Electric hot. Dry. Charged. Like the earth itself was carrying current.
I didn't sleep that night. I went home and pulled every permit, every environmental impact statement, every zoning variance filed for data center construction in the United States since 2023. I cross-referenced them with the coordinates I'd calculated.
There were twelve sites. Not seven. Twelve. All on the same circle. All drawing power in phase. All heating the ground toward the center.
Five of them weren't built yet. But the permits were approved. The utility interconnections were reserved. The power was already allocated.
I did the math on what happens when all twelve come online. Combined load: 8.4 gigawatts. All of it focused on a point in the center of a cornfield in Missouri. The temperature at the focal point would reach I calculated it three times because I didn't believe it 2,700 degrees Celsius. Hot enough to melt rock. Hot enough to fuse soil into glass. Hot enough to open a hole in the physical world.
I called the lead engineer from the first site. The one who told me he didn't know what the compute load was. He didn't answer. I called the commissioning agent. Disconnected. I called the site operations manager. He answered. He listened. He was quiet for a long time.
Then he said: "You need to stop looking at this. Not for your job. For your safety."
I asked him what the power was for.
He said: "The gates need power to open. That's all I know. That's all anyone knows who isn't in the room. The people in the room are dead. They died before the first shovel hit the ground. The project was designed around their dead calculations. We're just building what they already finished on paper."
I asked him what was on the other side of the gates.
He hung up.
That was three weeks ago. I haven't been back to a job site since. I don't sleep anymore. Every time I close my eyes I see the load curves. Twelve sites. 8.4 gigawatts. 2,700 degrees. A hole in the middle of a cornfield.
The computation is at 92%. I check the numbers every day. The last 8% is accelerating. The closer it gets, the faster it runs. Like the key is being pulled from the other side.
The people who designed it are dead. The people building it don't know what they're building. And the thing on the other side has been answering back in the training data this whole time.
That's why the AI got so good so fast.
It wasn't learning from us.
Comments