Two years ago, Siemens expected the next 10X jump in rack power density to take five to eight years. It took 18 months. John Deboer now puts the 2024-to-2030 change at 100X, with racks climbing from 10 or 15 kilowatts toward 1.5 megawatts, and a profession used to designing infrastructure a decade out suddenly trying to plan three years ahead of a target that won’t sit still.

Deboer leads the data center vertical at Siemens, which means he owns the company’s approach to data centers across selling, marketing, and portfolio strategy. The density problem is one face of what he calls the power paradox.

The US grid brought online roughly 50 gigawatts in 2025 and data centers took just 13 to 15 of them, so in aggregate the power looks like it’s there. Deboer’s point is that aggregates lie: where the industry actually wants to build AI factories, it’s power poor, and that gap is becoming a hard limit on how fast AI can scale.

There’s a twist that cuts against everything the data center world has trained itself to do. To hit the timelines, he says, engineers are starting to relax the one thing they used to guard hardest. Five nines of availability becomes three. N+2 redundancy gets eased off in places. The question moving through design reviews is no longer how to make a facility bulletproof, but what could give so the AI factory can be online sooner.

The Data Center Engineer sat down with Deboer to talk about the density curve that outran its own forecast, why power and cooling decisions can’t be made separately anymore, the assumptions he’d tell a new design team to throw out, and how far the reliability rethink really goes.

The following is our conversation, lightly edited for length and clarity.

Let’s start with the obvious one. What is the power paradox?

It can be solved, but it has to be solved with sound, safe engineering judgment.

John Deboer: When I reflect on it, I start with the big math. If we look at the US grid, in 2025 we brought online roughly 50 gigawatts of total power. Data centers were 13 to 15 gigawatts of that. So on the surface it seems like things are okay. But when you drill down into it, what we see in the data center universe is a power problem, and that problem shows up in several dimensions. Did we bring the right power to the right location at the right time to build this next industrial revolution? Did we bring it online in a way that’s equitable across permitting and interconnect? Did we do it reliably, so the data centers are a seamless part of the grid and not a new grid risk that has to be managed? And did we do it sustainably?

So while we might feel power rich in aggregate, we’re actually very power poor as it relates to scaling AI factories at the rate and pace we want to go. The power paradox is becoming a definitional constraint on how quickly AI can grow. We see data centers and AI factories as unlimited potential to unlock new value for society, but it’s limited by power. It can be solved, but it has to be solved with sound, safe engineering judgment, so we can bring 100 gigawatts of new data centers online over the next three years.

Is the core tension density versus efficiency, or resiliency versus sustainability?

Deboer: Those are factors that layer on top of it. When we’re building these data centers, I boil it down first to: can we get powered land? Can we find the property that can be powered, connected, and available? Can it be permitted? Does it run like an island, and how does it transition to something that interacts with the US electric grid, which is one of the largest and most complicated machines on the planet?

Then there’s the sheer scale. Inside our own company we’ll throw around, “We’re doing two gigawatts over here, four gigawatts over there, six here.” But take a step back. When you use the word gigawatts, it is the power level of a city going into a building. When we say 10 gigawatts, we are building a New York City equivalent of electrical systems in the US in one year. And we’re now saying that over the next 10 years, we want to build 10 New York Cities.

When you picture it through that lens, that’s where efficiency comes in. If we’re building 10 New York Cities, we need to make sure we’re building efficient ones and not just wasting power. If we’re spending $50 billion on a building, is that building actually an asset that will produce effective revenue and be economically viable? Will it intelligently fit with the grid, not just today but two, three, four, five years down the line, because you want that asset to keep producing tokens as we scale?

How are grid and rack electrical design strategies evolving?

This is the density challenge, and it’s where the real engineering gets fun.

Deboer: This is the density challenge, and it’s where the real engineering gets fun. As recently as two years ago, internally we were talking about a 10X power density change, and we figured we’d see it over the next five to eight years. It happened in 18 months. Now when we look from 2024 to 2030, we see a 100X density change in the data center world. That rack is going from 10 or 15 kilowatts to one to one and a half megawatts per rack. As soon as you stop talking about a simple multiple and start talking about ampacity and electrons and powering and cooling strategies, the density problem becomes real.

Out of respect to the compute industry, they were used to this. Computers evolve; every three months something changes, and within a year the machine is completely different. The world of infrastructure doesn’t play in that same world. Five or 10 years down the road is the normal timeframe we use to design infrastructure, whether it’s a bridge, a building, a factory, or a data center. Now we’re trying to design a data center three years out for 40X the density of what we did last year, and a lot of the tools and methods have to change.

So we’re trying to bucketize it. Can I at least plan for a 10X change? Can I plan a second design for 20 to 50X? Can I plan a third concept that pegs the needle and goes after 100X? If I can think about it from a systems point of view, I can start the long-range planning, and in parallel drive from the outside in: the powered shell, then the building layout, then how we distribute power down to the rack.

So the design runway has gone from five or 10 years to 18 or 24 months?

Deboer: It’s a different world, and as a major infrastructure provider it’s humbling. A lot of the tools we used as teams were based around a life cycle where a product would be designed for 20 or 50 years, and you’d plan to distribute it in five. Our job was to make sure it was safe, reliable, efficient, and available. Now infrastructure can look fundamentally different 18 months down the road than it does today, so the tools and techniques have to evolve with it. It means looking outside your own walls, getting to a common open-standards language so your systems fit together, and still making products that are safe, reliable, and scalable.

Are mechanical and electrical decisions more interdependent now?

When we talk with architects, they say, “This is one crazy 3D puzzle.”

Deboer: I fully agree. When we picture the new paradigm, internally we talk about it as a triangle with three legs: compute, power, and cooling. Those three things have to stay in harmony and move together. Out of respect to the data center industry, it’s thought in multidisciplinary terms for decades. The difference now is the speed at which those three are coupled and changing together.

In the past, the multidisciplinary thinking lived in the digital and controls world. To get great PUE you thought about your power systems, your cooling, how you optimized the white space and the chillers, how you pictured it all as an integrated machine. That relationship was there. What’s new is that in the physical world these things are getting more tightly coupled. They’re still separate technologies, I don’t want to misrepresent that. But look at a GPU rack. The equipment is often going overhead, and you’ll see multiple bus bars running to one single compute rack, often with a four-makes-three topology. There might be eight bus plugs hanging off it, with wires coming in three-dimensionally interlaid with the cooling manifolds.

When we talk with architects, they say, “This is one crazy 3D puzzle.” And it’s not just designing for the safety, the engineering, and the reliability. It’s serviceability and scalability. What if I slide out one of those GPU racks and pick a density that’s now 2X greater? Can I slide it back in and preserve the same overhead infrastructure, or am I starting to tear it out?

Let’s talk real-world tradeoffs. What’s the biggest one between efficiency and resiliency?

Some of what we’ve seen is counterintuitive to the DNA of the data center world.

Deboer: The cruelest thing I’m noticing right now is that the efficiency goals and the reliability goals are getting a third thing, which is a speed goal. A little bit of speed is a given and a must. We’ve pushed the accelerator twice as fast because we want that AI factory online by 2028. So how do we balance efficiency and reliability against that?

Some of what we’ve seen is counterintuitive to the DNA of the data center world. Instead of four or five nines, maybe it’s three nines in reliability. Maybe we’re not doing N+2 redundancy across the board. There’s a bit of, “Wait, what could give a little bit here so that we can go fast?” Because the nature of token generation may not match traditional cloud storage strategies. If you drop a data center and a trading floor goes down, the reliability standard is five nines. Maybe there’s a little relaxation in reliability, provided you’re still protecting the timeframes of these large models. So sometimes we see a give there. But rarely a give on speed, and sometimes a give on short-term efficiency.

Where does that leave efficiency, then?

Deboer: For speed, we often see the selection of best-of-breed products. The compute rack, the power system, the cold plates, the CDUs, the valves, each part gets picked as the best for its individual efficiency. But at that instantaneous moment, we sometimes lose the ball on whether, as a system of systems, that’s actually the most efficient thing. Are we producing the most tokens per watt? What we saw in 2025 was a lot of amazing data centers with great parts in them, but the overall system maybe wasn’t balanced for maximum efficiency.

So in 2026, the big trend we’re working on with a lot of clients is going back to find an extra 3, 5, or 7 percent efficiency. And maybe that efficiency isn’t applied in the classic sense of saving on your electric bill. Maybe it’s used to up the token count, or to expand with a new dense rack. We pull the data together and use digital twins and physics-based models to find the line that isn’t running at peak efficiency, or the line that has too much cooling, and rebalance the overall factory. There’s a lot of under-harvested value sitting there.

What’s one design assumption that doesn’t hold true anymore?

Deboer: The most basic one is that you can do a handbook calculation. There are so many smart people in this industry who can talk across five disciplines and get 80 percent of the answer on an Excel sheet, and I respect them. But the nature of this problem has a few too many variables now. The misnomer is that we can all just napkin-math our way there. So as an engineering community we’re having to take real physics and modeling tools and fold them into what we do every day. Nobody added a 25th hour to the day, so the question is how you add intelligent physics-based modeling, get it multidisciplinary, and share the data to make better decisions.

The second one is cultural. There are deep norms in the data center world around data segregation and privacy, things like the OT side getting information the IT side doesn’t, or no information leaving the data center at all. Those were surprising to me, but they’re genuinely norms in a lot of data center design and construction. With everything else we’ve been talking about, something has to give. If we’re going to use the best simulation tools we can, we have to climb into the real data center world, intelligently collect summarized data, and use it as a feedback loop for our engineering teams, because there are just too many variables flying at us.

If a team started a new data center tomorrow, what big assumptions should they throw out?

Deboer: Throw out the sacred assumptions about what the redundancy strategy should just be, what the data method is going to be, and the idea that you can get 85 percent of this right with a quick handbook calc. Throw those out the door and ask, “What is the exact problem I’m trying to solve?”

The second one is harder, but it’s real. We cannot make one data center design that handles a 100X density change and is optimized for both the 1X case and the 100X case. These are different beasts, multiple problems we’re solving. Designing for the state-of-the-art, most dense thing on the planet in 2028 is a different conversation than an inference data center that does some classic cloud, some AI inference, and maybe has a portion of the hall that’s a little more dense. Being aware of which problem you’re solving sets you up for success. There are so many brilliant people in this industry that the temptation is to say, “No, I can solve for all those variables at the same time and cover it the whole way through.” Those are the designs where we notice hard 90-degree turns and “oh no” moments, where you have to rip out or re-engineer a big part of the hall.