In September 2024, Microsoft signed a 20-year power purchase agreement to support the restart of Unit 1 at Three Mile Island — a nuclear plant shut down for 50 years after the 1979 accident. The reason for the restart is singular: to power AI data centers.
On the same timeline, Google and Amazon signed small nuclear reactor deals; Elon Musk secured dozens of gas generators in Tennessee in one sweep; OpenAI's Stargate project locked in nearly half a trillion dollars and more than a dozen gigawatts of data centers; and Jensen Huang keeps repeating one line in public: "The bottleneck of AI is electricity — the end of compute is power."
That AI's expansion has shifted from a compute race into an energy race is now consensus. But push the energy race further and what you hit is not a resource question — it is a series of physical engineering questions: cable thickness, grid approval timelines, cooling efficiency, the world's annual copper output. The pace of the physical world is now setting how fast AI can run.
The energy constraint on AI data centers can be broken down along two dimensions: finding power (whether you can secure enough electricity) and using power (how to deliver it to the GPUs efficiently and stably once you have it). Both ultimately collapse into the same problem: the physical world's constraint on the digital world's expansion is shifting from soft to hard.
Power Density Explodes: An Order-of-Magnitude Leap
Data centers are nothing new. In the cloud computing era, a standard rack drew 10 to 30 kilowatts, and almost nothing exceeded 30. AI changed that number.
Nvidia's H100 rack jumped to 40 kilowatts — double the traditional cloud baseline. The next-generation GB200 jumps to 120–140 kilowatts, tripling again. The generation after that — the Rubin Ultra architecture slated for mass production in 2027 — reaches 600 kilowatts to 1 megawatt per rack: tens to over a hundred times the traditional cloud baseline.
A typical AI data center is gigawatt-scale. One gigawatt is roughly the electricity consumption of a medium-sized city of 8 million people. Stacked together, these numbers turn a data center from a server room into, as the source video puts it, "a giant monster that devours land, water, electricity and chips — and excretes tokens."
Power-related infrastructure and electricity bills already account for one-fifth to one-quarter of a data center's full life-cycle cost. Electricity has moved from being a cost line to being a factor of production — if power can't keep up, even the most advanced GPUs can only sit there depreciating.
Finding Power: A Structural Divide Between the US and China
China and the United States face entirely different conditions when it comes to finding power.
China's grid has national-level top-down design and dispatch. The "Eastern Data, Western Computing" project laid out eight backbone nodes, and data centers deployed near those nodes are, in aggregate terms, reasonably well supplied.
America's situation is far more complicated. First, the grid is aging and upgrades are slow. The deeper problem is that US grids are managed state by state — a gigawatt-scale data center is the equivalent of a small-to-medium city, and wherever it lands, local residents push back. Winning grid-connection approval from the government means waiting in a queue for years.
AI giants that cannot wait have no choice but to generate their own power. Their options form a menu arranged by time scale:
- Long-term solution: nuclear. Stable, zero-carbon, available 24 hours a day. But a new reactor takes 5 to 10 years from approval to first power. Microsoft, Google, Amazon and Meta are all signing long-term power purchase agreements, betting on exactly that time scale.
- Medium-term bridge: natural gas. Fast — usable within months. The price is emissions, and continuous warnings from environmental regulators. Musk has secured large permits for gas turbines in Mississippi and Tennessee.
- Short-term supplement: wind, solar plus storage. Unstable — the sun disappears without notice, the wind fluctuates. For a data center that demands extreme stability, the intermittency of green power is itself the problem.
- Going offshore to detour around regulation: build where rules are looser — OpenAI is negotiating a large data center in the United Arab Emirates; Amazon and Oracle are building in Spain and Thailand. Going further still, Musk and Google have considered moving data centers into space.
"When it comes to finding power, the giants have truly played this game to the limit. Sky or ground — each one dares to dream bigger than the last."
Xiaolin Shuo ("Little Lin Says") is a popular Chinese financial explainer channel by Lin Wei, whose video on AI's energy war is the source of this essay.
Using Power: An On-Site Deconstruction of a Huawei Data Center
Finding power is only half the story. Once the electricity is secured, how to deliver it to the GPUs efficiently and stably is the other half. Xiaolin Shuo went inside a Huawei AI data center and walked through the full chain, from the power supply pods to the liquid-cooled racks.
Three Steps of Power Delivery
Electricity travels from the grid to the servers through three layers of processing: incoming feeders — high-voltage grid power stepped down to 380 volts by transformers; the UPS (uninterruptible power supply) — converting unstable grid power into the stable power servers can use; and outgoing feeders — distributing power to every rack and every server.
The critical piece of equipment is the UPS. Its working principle resembles a computer power supply, but the requirements are orders of magnitude higher: server capacitors can only ride through milliseconds of outage, so the UPS must complete its switchover within a few milliseconds. Huawei's solution is modular design — the UPS is assembled from many small modules, each of them hot-swappable, so maintenance means pulling out the old module and inserting a new one, with no shutdown required.
Outside the UPS sit lithium batteries (good for 10 to 15 minutes), and outside those, diesel generators can be connected (good for far longer). Every tier has redundant backup.
At the 300-kilowatt rack level, a single cable approaches the thickness of a milk-powder can — and several such cables must enter one rack simultaneously. Physically, they do not fit. This is the first "physical limit" AI data centers have run into.
Liquid Cooling
A 300-kilowatt rack is the thermal equivalent of one to two hundred electric stoves running at once. If the heat cannot escape, the servers simply burn up.
Huawei's answer is liquid cooling: chilled water flows through supply pipes into micro-tubes inside the servers — like capillaries — absorbs the GPUs' heat, and flows out as hot water. Liquid cooling carries away roughly 95% of the heat, with air cooling handling the rest. The hot water converges at a thermal management unit (TMU), which monitors the temperature, flow rate, velocity and pressure of every loop in real time and adjusts the water flow dynamically.
The TMU does more than manage heat exchange — it also performs predictive maintenance. A liquid cooling system's two worst fears are leaks and blockages: a leak can cause a short circuit, a blockage traps heat, and both can destroy GPUs. Huawei's approach is to train AI models on operations data so that even micro-leaks trigger an alarm before they happen.
The Cables Are Too Thick to Fit: A Physical Limit Exposed
Power equals voltage times current. To raise power, you raise voltage or you raise current. But the larger the current, the thicker the cable must be. At the 300-kilowatt rack level, cables grow to milk-powder-can thickness — and a rack only has so much space.
The direction of the solution is high-voltage DC or 800-volt supply: higher voltage means lower current for the same power, which means thinner cables. The catch is that heavy electrical engineering is not like information technology: the grid frequency is 50 hertz, and moving to an 800-volt standard means every piece of distribution equipment, every switch and every component must be adapted to the new standard. This is not a technology problem — it is an ecosystem adaptation problem.
Jensen Huang has publicly endorsed this direction, and Huawei is pursuing it too. Both sides agree high-voltage DC is where things are heading, but the route is "multiple architectures coexisting" — not a single overnight switch.
From Consumer to Grid Partner: Grid-Forming Capability
AI data centers have one more distinctive power characteristic — the GPUs' "short temper."
During training, tens of thousands of cards spin up simultaneously and the load surges to the ceiling; when a training phase ends and checkpoints must be saved, the load drops off a cliff. For the grid, that means a load that is alternately enormous and gone. At gigawatt scale, such swings are enough to affect regional grid stability. If grid dispatch cannot keep up, generators can trip — potentially cascading into large-scale blackouts.
The traditional answer is to let the UPS absorb the shocks. Huawei instead proposes a three-tier capability ladder:
- Grid adaptation: even when grid quality is mediocre, the data center keeps running on its UPS.
- Grid support: when a local grid fault occurs, the data center must not simply disconnect — it must reconnect within 0.5 to 1 second, or load and supply fall out of balance.
- Grid-forming: the data center actively supplies reactive power back to the grid — where the grid once fed you, now you help the grid hold its voltage steady. From consumer to partner.
The technical term is source-grid-load-storage — power source, grid, load and storage working as one coordinated system. Huawei's idea is to use this framework to transform the data center from a "power tiger" into a "voltage regulator."
The Physical World Is Holding AI Back
Back to the opening question: how did AI become an energy war?
Look one layer deeper and the answer is a collision between the physical and digital worlds. For decades, the internet grew mainly on code and software, with little friction from the physical layer. Today's AI, by contrast, is literally burning water, electricity and silicon into tokens. Its expansion speed must obey the pace of grid construction, nuclear approval timelines, transformer delivery schedules and the world's annual copper output.
By 2026, the frontier AI companies had collectively returned to energy and infrastructure — industries that look the plainest and least glamorous. Moving data centers into space sounds like science fiction, but it is precisely a reflection of the hard physical constraints here on Earth.
"AI is a textbook case of the physical world holding back the speed of the digital world. Software used to run faster than hardware; now the hardware's physical limit is written on the wires — the cables are too thick to fit."