From Chips to Kilovolts: The Rise of the 10-Gigawatt Data Center
Ten years ago, building a 100-megawatt data center was a big deal. You'd put up a few buildings, hook into the local power grid, and call it a day. Now, companies are planning 10-gigawatt sites—that's a hundred times bigger. To give you an idea, 10 gigawatts is roughly what all of New York City pulls on a mild Tuesday afternoon. We're not talking about a data center anymore; we're talking about a city-scale industrial complex. And with that shift, the biggest bottleneck for AI has moved from how many chips you can buy to how much power you can get. It's a whole new game.
The megawatt-to-gigawatt mindset shift
I remember touring a data center in Oregon back in 2014. The facility was maybe 20 megawatts, and the tour guide kept bragging about how they had a dedicated substation. Back then, that was cutting-edge. Fast forward to today, and Softbank is planning a 10-gigawatt site in Ohio. That's not just a bigger building—it's a fundamental change in how we think about infrastructure. Microsoft is doing similar things in the MISO territory. These sites are so large that they stop being customers of the grid and become part of the grid itself.
The old metric, PUE (Power Usage Effectiveness), still matters for internal efficiency, but it's getting overshadowed by something called the interconnection queue. In many parts of North America, if you want to hook a new large load to the grid, you might wait over five years. Five years! Developers can't sit around that long. So they're starting to build their own high-voltage transmission lines and substations. The days of plug-and-play data centers are over. Back in the day, you picked a location based on tax breaks and fiber routes. Now, you pick a location based on where the high-voltage lines are underused and where regulators won't freak out about your power draw. Ohio has become a hotspot because it has all these old transmission lines from when factories were booming. Those factories closed, but the power grid stayed, and now tech companies are breathing new life into it.
The new gatekeepers: grid operators
Getting the power is half the battle. The other half is figuring out how to deliver it without crashing the grid. Organizations like MISO and PJM manage the flow of electricity across state lines. They make sure that when your data center turns on, the lights don't dim in a nearby town. Microsoft has been partnering with MISO to get a better view of congestion and plan for the long term. These grid operators use crazy weather models and load-balancing algorithms to keep the AC frequency stable. A 10-gigawatt data center uses power at a flat rate—24/7, no ups and downs. That's good for utilities because it's steady revenue, but bad because it offers zero flexibility. Unlike a neighborhood where demand drops at night, an AI training cluster runs flat out all the time.
Because of that, developers are looking at "behind-the-meter" solutions: building their own power source right next to the data center. If you own both the power plant and the servers, you don't have to worry about price spikes or utility delays. You remove the biggest risk to uptime. It's like having your own private generator, but on a massive scale.
The physical limits of heat and transmission
Now, let's talk about the engineering challenges. Moving 10 gigawatts of electricity requires the biggest transmission lines we have—765 kV. Those lines are a nightmare to permit and expensive to build. And once the power gets to the site, you have to deal with the heat. 10 gigawatts in equals 10 gigawatts of heat out. You can't cool that with air. I mean, you could try, but you'd need fans the size of jet engines. Even the best HVAC systems can't move enough air. So everyone is moving to liquid cooling—either direct-to-chip cold plates or full immersion. These systems need massive amounts of water and fancy chemical treatment plants to keep the loops clean.
But there's another problem: distance. When your data center campus spans miles, the time it takes for data to travel between racks becomes a factor. Engineers have to design networks that account for the speed of light over kilometers of cable. And they have to make sure the high-voltage power lines don't interfere with the fiber optics. It's a whole new level of complexity. I talked to a network engineer who said they now spend more time on electromagnetic interference than on routing protocols.
Power as the ultimate competitive moat
In the early days of cloud, the moat was the software. Then it became the custom chips. Now, the moat is the energy contract. If you lock up a 20-year deal for 5 gigawatts of nuclear power, you have a structural advantage that no algorithm can beat. It's like going back to industrial-era thinking—whoever controls the power wins. That's why we see companies racing to restart old nuclear plants or buy into Small Modular Reactors (SMRs). If a tech firm can put its own reactors on site, it can bypass the grid entirely. They create a "sovereign compute zone" where the only limit is how many chips they can get.
But this concentration of energy and compute creates new risks. If the AI race comes down to who builds the biggest power moat, we might end up with a handful of massive hubs controlling everything. Those sites become critical national infrastructure. They need their own security, logistics, and emergency plans. The cloud is no longer something abstract—it's grounded in copper, steel, and uranium.
The end of the virtualized era
For years, the tech industry tried to pretend hardware didn't matter. We talked about serverless and the cloud as if they were magic. But a 10-gigawatt data center shatters that illusion. Every AI output has a direct cost in electricity and water. The people building these sites aren't just DevOps engineers anymore; they're negotiating with senators and studying river flows. The most successful software companies of the next decade will probably be the ones that are best at heavy civil engineering and utility-scale power management.
So, what does all this mean? We're heading toward a world where compute power is concentrated in a few places. Does that make us fragile—one grid failure away from losing the world's smartest AI? Or will it force the industry to build more resilient, decentralized power systems? Maybe both. Either way, the days of pretending infrastructure is boring are over. It's the new battleground, and it runs on kilovolts.
Let me tell you a story. I have a friend who works on one of these mega-sites. He told me that last year, they had to buy a small substation from a local utility. The negotiation took nine months. Nine months for a single transformer. That's where we are now. The bottleneck isn't innovation—it's permitting. And until we figure out how to speed that up, every new AI model will be limited by how fast we can build power lines.