When we talked about cooling in the previous gate, one line came up: AI’s real ceiling is shifting from “the chip” toward “cooling and power.” This gate is where we lay out the “power” part clearly. When every rack of GPUs acts like a power-devouring monster, the electricity an entire AI data center needs has grown big enough to put utilities and governments on alert.
This piece breaks down, in plain English, what compute is, just how much power an AI data center uses, how much the cloud giants are spending, and why power is becoming a hard limit on AI’s expansion. This is the deep-dive version of Gate 7, “data centers and power,” in The AI Hardware Supply Chain, End to End.
What Compute and AI Data Centers Are
Let’s plain-talk two terms first. Compute is “the ability to do calculations,” commonly measured in FLOPS (roughly, how many number-crunching operations a computer can do per second). AI training and inference need to compute staggering amounts, so they cluster tens of thousands of GPUs together to provide compute.
An AI data center is, at its core, a large facility that integrates compute, memory, high-speed networking, liquid cooling, and power gear in one place. Its biggest difference from a regular machine room is power density, cooling, and load swings: a traditional rack might draw a few kilowatts, while an AI rack routinely draws over a hundred kilowatts, and AI computing’s power use swings sharply within short windows. The International Energy Agency (IEA) estimates that by 2027 a single AI rack’s peak draw could be equivalent to about 65 households.
Just How Much Power?
Let’s go straight to the scale. By the IEA’s estimate for a 100MW-class hyperscale AI data center, annual power use can be equivalent to about 100,000 households; the largest sites under construction can run an order of magnitude higher still.
Pull the lens back to the whole world: under the IEA’s Base Case, data centers’ share of global electricity is climbing from about 1.5% in 2024 toward about 3% by 2030. That doesn’t sound like much, but the growth is fierce. In 2025 alone, global data center electricity use rose about 17% year over year, and AI-focused data centers rose about 50%. And this demand is highly concentrated: the US alone accounts for nearly 45% of global data center electricity use, so the strain on local grids runs far above the global average.
Core-Data Snapshot
The numbers below help you grasp the scale of AI’s power use. Most are estimates from international agencies, so reading the trend is more practical than fixating on absolute values.
| Topic | Data | Time / Nature |
|---|---|---|
| Data centers’ share of global electricity | About 1.5% in 2024 → about 3% in 2030 | IEA estimate |
| 2025 global data center electricity growth | About 17% (AI-focused about 50%) | IEA |
| One large AI data center’s power use | Roughly equal to 100,000 households; biggest sites reach the million-household range | IEA framing |
| US share of global data center electricity | About 45% | 2024, IEA |
| Cloud giants’ 2026 capex (combined) | On track for about $610-725 billion | 2026E, multiple estimates |
| Single AI rack’s peak draw (2027) | Roughly equal to 65 households | IEA estimate |
Where the Money Goes: The Cloud Giants’ Capex
The flip side of AI data centers eating power is eating money.
The capex (capital expenditure — the long-term investment in building data centers and buying servers and power gear) of cloud giants like Amazon, Microsoft, Google, and Meta is surging. Multiple firms estimate a combined 2026 figure on track for $610 billion to over $725 billion, roughly triple what it was two years ago. In Bloomberg’s estimate, Amazon alone could spend up to $200 billion in 2026, with Google and Microsoft at about $190 billion each.
How big is that scale? Roughly compared against US nominal GDP, $700-plus billion is around 2% of it. A substantial chunk of that will flow into GPUs, memory, advanced packaging, liquid cooling, power gear, and facility construction — effectively pulling the entire AI hardware supply chain up along with it.
Why Power Is Becoming the Ceiling
In the past, when people talked about AI’s expansion, the bottleneck was “can you even get GPUs.” The wind has shifted.
The IEA puts it bluntly: AI data centers’ progress is bottlenecked at generators, transformers, chips, IT components, grid connection, and approvals. Building a data center takes about two or three years, but laying a grid line thick enough — and waiting on transformers and interconnection permits — often takes far longer. In other words, GPUs still matter, but “can you connect enough stable power” is becoming an equally — even more — binding constraint.
That’s also why the big players have started treating “power” as a resource to grab: signing long-term power purchase agreements, building their own battery storage, and locking up generator and transformer capacity. Whoever solves power supply first is the one who gets to talk about expanding compute.
Nuclear and SMRs: A Long-Term Fix, Not Yet Reliable Short-Term
Power is so tight that even nuclear has been put on the table.
Amazon, Google, Microsoft, and Meta have all signed nuclear-related deals over the past couple of years. Amazon is investing in X-energy and planning small modular reactors (SMRs); Google’s demonstration reactor with Kairos Power broke ground in 2026; Microsoft signed on to support the restart of an existing nuclear plant in Pennsylvania; and Meta has signed long-term clean-energy deals with the likes of TerraPower and Oklo.
But realistically, most of these are long-game positioning. Most SMRs won’t deliver real power until after 2030, and restarting existing plants takes time too. In the near term, AI data centers mainly lean on existing nuclear, natural gas, and renewable power purchases, plus battery storage, to hold up. It’s more realistic to treat SMRs as a post-2030 long-term backstop than to assume they’ll feed data centers tomorrow.
The Power Architecture Is Changing Too, and Taiwan Has a Part
Power isn’t just about “enough or not” — it’s also about “how you deliver it.”
As rack power draw pushes from a hundred-plus kilowatts toward megawatt class, the traditional power architecture starts to strain on efficiency and copper losses. NVIDIA is therefore pushing an 800-volt DC (800 VDC) data center power architecture, aimed at supporting racks from 100 kilowatts to megawatt class. Taiwan has a part here too: high-efficiency power supplies, 800 VDC power racks, uninterruptible power systems and storage, busbars and cabling — Taiwanese firms hold relevant supply-chain roles across these links, and NVIDIA’s 800 VDC ecosystem list also names Taiwan-related firms such as Delta, Lite-On, and Richtek (being listed means part of a partner ecosystem, not that orders have shipped or that it’s any kind of investment guarantee). This only describes public supply-chain roles; it does not compile beneficiary stocks or buy/sell advice.
Key Takeaways for This Gate
After this gate, first remember the scale: one large AI data center uses as much power as 100,000 households, the global share is climbing toward 3%, and the US is under the most pressure; the cloud giants’ combined 2026 capex runs to the $700-billion class.
The more crucial judgment is this: AI’s ceiling is shifting from “the chip” to “power.” GPUs still matter, but connecting power, supplying it stably, and getting transformers and interconnection permits are becoming an equally — even more — binding constraint. Nuclear and SMRs are a long-term fix, mostly delivering real power only after 2030; in the near term it’s still existing grids, natural gas, renewables, and storage.
To see how all that power finally gets carried away as heat, and what the heat-generating beasts look like, check out What Is Liquid Cooling and What Is Blackwell; to see how all eight gates of the chain string together, head back to the supply-chain overview.