INTRODUCTION
This article is Part II of a three-part series on AI Data Centers. These articles are a response to a reader who asked these three questions:
1. Which public issues are impacting AI Data Centers?
2. A few of us believe their expansion is limited by physics. Is that so?
3. And how long will this last? Is an AI bubble forming?
In response, I am answering in three parts, as follows:
· Part II: Physical Limits
· Part III: AI Bubble
Listen to it first and then read it afterward!
COMPLIANCE SOLUTIONS
AI POLICY PROGRAM FOR MORTGAGE BANKING™
Policies and procedures for artificial intelligence related to mortgage banking compliance administration, mortgage loan origination compliance, mortgage servicing compliance, and industry best practices—including, but not limited to, applicable federal and state banking and consumer lending laws. Currently, it consists of nine policies, with more planned.
AI FOR MORTGAGE LOAN ORIGINATION - SALES MANUAL
The Sales Manual provides 9 chapters covering the full AI-powered origination funnel in a 90-day implementation roadmap, with clear team roles and fair lending, FCRA, and advertising compliance guidance built into every chapter. Sample AI prompts, an ROI tracking framework, and a compliance glossary.
RESPONSE
The physics: why you can't just build faster
I concentrated in math and science in college, specializing in symbolic mathematical logic, and I have peer-reviewed papers in that field for scholarly journals. Computer science is not my specialty, but I studied it, and for over fifty years I have taken part in science and physics forums, including on artificial intelligence (AI). Today, I contribute regularly to two math groups whose members include leading AI experts and academics.
Here, I want to give a layman's view of the physics of AI Data Centers, because the physical limits shape everything else. Parts of this article are technical. If that's not for you, stay tuned to Part III, coming soon, where I discuss whether there is an AI Bubble.
In my view, the binding limit on AI Data Centers is no longer silicon. It is the physics of moving energy in and heat out, plus the slow-to-build infrastructure that does both. The International Energy Agency (IEA) projects that global data center electricity use will more than double, from 415 terawatt-hours (TWh) in 2024 to 945 TWh by 2030. One TWh is a trillion watts running for one hour, or roughly what 100,000 American homes use in a year.
Every watt becomes heat
Let's begin with the law of conservation of energy: energy is never destroyed, only converted. It is unforgiving! Nearly all the electricity fed to a Graphics Processing Unit (GPU) leaves the building as heat.
A GPU is a chip that performs thousands of calculations at the same time. It was built to draw graphics, images, and video, but that same ability to process huge amounts of data at once now powers AI, machine learning, and scientific research.
The largest AI Data Centers are gigawatt-scale facilities, called a GW Campus in AI Speak. Each uses one or more gigawatts (1,000+ megawatts) of electricity to train and run frontier models, the most capable AI models available at any given time.
Now, please follow me, as I bullet the physics:
· A 1 GW campus is, in thermodynamic terms, a 1 GW heater that must dump its heat somewhere. Thermodynamics is the branch of physics that deals with heat, work, temperature, and energy.
· The cheapest place to dump it is into evaporating water. Evaporating one liter of water absorbs about 2.26 megajoules (MJ) of heat. A joule is the standard unit of energy; a megajoule is a million joules.
· One kilowatt-hour (kWh) of electricity equals 3.6 MJ, so getting rid of it purely by evaporation takes about 1.6 liters of water.
· For a 1 GW campus running all day, that is roughly 38 million liters, or about 10 million gallons, per day.
· Dry cooling avoids using water, but it needs more fan power, and it works worst on hot days, exactly when the grid is most stressed.
That arithmetic is why protesters fight to protect their water supply and an increase in electricity costs.
Air has run out of road
The same volume of water holds about 3,500 times more heat than air. That is why air cooling tops out at roughly 40–50 kilowatts (kW) per rack: beyond that, the fans use more power than the cooling they add. That limit has been called "the thermal wall nobody saw coming."
AI racks have blown far past it: about 40 kW for Nvidia's Hopper, 120 kW for Blackwell, and about 600 kW for Nvidia's Rubin Ultra "Kyber" rack, due in 2027, with 1-megawatt (MW) racks on the roadmap.
So, liquid cooling is now mandatory. A hall built for 2024 racks cannot host 2027 racks without rebuilding its power, its cooling, and even its floors, which must carry much heavier racks.
Ohm's law sets the voltage
Think of electricity in a wire like water in a pipe. Voltage is the pressure, current (measured in amps) is the flow, and resistance is how narrow the pipe is. Ohm's Law ties the three together, and engineers use it every day to make sure electronics get the power they need without overheating.
Two simple rules matter here. First, power equals voltage times current. Second, the heat a wire wastes grows with the square of the current: double the current, and the wasted heat quadruples.
So, to deliver a lot of power, you want high voltage and low current. At the traditional 54-volt rack supply, a 1 MW rack would draw about 18,500 amps, far too much for practical copper wiring. Raising the supply to 800 volts of direct current (DC) cuts the current about 15-fold and the wasted heat more than 200-fold.
Nvidia and its suppliers are already making that move, but electrical codes and arc-flash safety standards for high-voltage DC lag the hardware by years. Until they catch up, each site needs its own approvals. This is obviously a less-than-ideal situation!
The grid must balance every second
Electricity is not stored at scale. Power plants must match demand second by second, and the wires between them and the users must be sized for the peak.
Again, let's bullet it!
· Data centers take 2–3 years to build, but grid and generation infrastructure take 4–8 years, according to the IEA, as cited in the prospectus for ERock's Initial Public Offering in June.
· Power transformers take about 128 weeks (about 2½ years) to deliver, and generator step-up transformers 144 weeks. In Texas, for instance, the grid operator's (ERCOT's) queue of large new loads jumped from 63 GW to 226 GW in about a year.
· Indeed, of roughly 16 GW of US data center capacity targeted for 2026, only about 5 GW entered active construction.
· Importantly, AI loads are hard on the grid. Training jobs make power use swing sharply, and data centers that drop offline together during a voltage dip can destabilize the whole system. Grid regulators have flagged this risk since 2024.
The speed of light keeps clusters together
Light in optical fiber travels about 200 kilometers per millisecond. That sounds fast, but training a frontier model requires thousands of GPUs to stay in constant sync, so every extra kilometer slows every step.
As a result, builders favor one giant campus over many small ones, which concentrates the power, water, and noise burden on a single community.
Chips no longer get cooler as they shrink - That's a big problem!
When I was in school, we learned the physics of Dennard scaling: as transistors shrank, their voltage and current shrank with them, so a chip's power per square inch stayed constant. Smaller chips meant more computing for the same power.
That ended around 2006. Chips still get more efficient per calculation, but today, more computing generally means more watts.
But – and it's a big But – there is no hard physical wall nearby!
The Landauer limit sets the absolute minimum energy needed to erase one bit of information. Information is physical: resetting a bit to a blank 0 or 1 reduces disorder, and that must release a tiny, unavoidable burst of heat.
Today's chips use hundreds of millions of times more energy per operation than that minimum. So physics leaves enormous room for more efficient computing. What stands in the way is the engineering: each jump in rack power forces a rebuild of the electrical room, the cooling plant, and even the floor.
So, the limits are engineering and infrastructure, not fundamental physics.
CONCLUSION
Physics does not put a ceiling on AI Data Centers. It sets a speed limit. Every limit I have described comes down to three facts:
· Energy in equals heat out. Every gigawatt of computing is a gigawatt of heat that must go into the air or into water. More AI means more cooling, and more cooling means more water or more power.
· Power must be delivered and balanced in real time. Higher voltages ease the wiring inside a rack, but the grid outside must still generate, carry, and balance every watt, and its transformers, turbines, and power lines take years to build.
· Distance costs time. Because frontier training needs thousands of chips working in lockstep, the speed of light pulls everything onto one site, concentrating the load on one grid and one community.
None of these is a law that says "stop." The true physical floor, the Landauer limit, sits hundreds of millions of times below today's chips, and efficiency will keep improving, likely faster than the grid can grow. And that's a huge limit in itself!
But efficiency has not slowed demand. As computing gets cheaper per operation, we use more of it; economists call this the Jevons paradox. So, for the rest of this decade, I expect the pace of AI Data Center growth to be set less by chip roadmaps than by how fast we can pour concrete, build transformers, string transmission lines, and win local permits.
In short, AI Data Centers are limited by physics in two ways: energy and heat can't be wished away, and the infrastructure that handles them is slow to build. The gap between what investors want built and what physics and the grid can deliver is where Part III begins. Is the money racing ahead of what can be built and used? And is an AI Bubble forming?
__________________________
This article, AI Data Centers: Physical Limits – Part II, published on October 8, 2026, is authored by Jonathan Foxx, PhD, MBA, the Chairman & Managing Director of Lenders Compliance Group, founded in 2006, the first and only full-service mortgage risk management firm in the United States, specializing exclusively in residential mortgage compliance.