In the previous gate on cooling, we said AI’s ceiling is moving from “the chip” toward “cooling and power.” This gate explains the power part. The electricity a whole AI data center needs is now big enough to get the attention of utilities and governments.
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
Compute is “the ability to do calculations,” commonly measured in FLOPS (roughly, how many number calculations per second). AI training and inference need a huge amount of calculation, so thousands of GPUs have to be grouped together.
An AI data center is a large facility that brings together compute, memory, networking, cooling, and power. It differs from a regular server room in three ways: power density (a regular rack draws a few kilowatts, an AI rack often over a hundred), cooling (it needs liquid cooling), and power use that swings sharply within a short time. The International Energy Agency (IEA) estimates that by 2027 a single AI rack’s peak draw could equal 65 households.
Just How Much Power?
The IEA estimates that a large 100MW-class AI data center uses about as much electricity a year as 100,000 households. New-generation sites are dozens of times bigger: according to CNBC, Meta’s Hyperion site in Louisiana is planned at 5 GW, and OpenAI says the PORTS-Pike site in Ohio it has joined is planned for about 8 GW of IT power.
At the global level, here are the IEA’s April 2026 estimates:
| Metric | Value |
|---|---|
| Global data center electricity use | 485 TWh in 2025 → about 950 TWh in 2030 (about double) |
| Share of global electricity in 2030 | About 3% |
| Electricity growth in 2025 | All data centers +17%, AI-focused data centers +50% |
3% doesn’t sound like much, but data centers cluster in a few places, so the strain on local grids is far greater than the global average.
Where the Money Goes: The Cloud Giants’ Capex
The flip side of eating power is eating money. Capital expenditure (capex) is the long-term investment in building data centers and buying servers and power gear. Here is the full-year guidance the four cloud giants gave in July 2026:
| Company | 2026 capex guidance | Source |
|---|---|---|
| Amazon | About $220 billion | Official earnings call (2026-07-30) |
| Alphabet (Google) | $195–205 billion | Company guidance, according to SiliconANGLE (2026-07-22) |
| Microsoft | About $175 billion (calendar 2026, after an accounting adjustment; previously $190 billion) | According to CFO Dive (2026-07-29 earnings call) |
| Meta | $130–145 billion | Official results (2026-07-29) |
| Guidance total | About $720–745 billion | Cash and lease bases differ by company |
Much of this money flows into GPUs, memory, advanced packaging, liquid cooling, power gear, and facility construction, which lifts the whole AI hardware supply chain along with it.
Why Power Is Becoming the Ceiling
AI’s bottleneck used to be “can you get GPUs.” Now it is also “can you get power.” The IEA’s 2026 report says supply of gas turbines, transformers, advanced chips, and IT parts is tight, and that a flood of data center applications has clogged approvals and grid connections. Building a data center takes about two or three years, but waiting for transformers and an interconnection permit often takes longer.
So the big players are racing for power: signing long-term power purchase agreements, building natural gas generation on site, and adding battery storage. Only whoever solves power first can talk about expanding compute.
Nuclear and SMRs: A Long-Term Fix, Not Yet Reliable Short-Term
All four cloud giants have signed nuclear-related deals, but their progress varies a lot:
- Microsoft: signed a 20-year power purchase agreement backing the restart of the Crane plant in Pennsylvania (formerly Three Mile Island Unit 1); the operator, Constellation, targets 2027.
- Google: the Hermes 2 demonstration reactor with Kairos Power broke ground in April 2026 and will put up to 50 MW on the grid.
- Meta: signed with Vistra, TerraPower, and Oklo, aiming to support up to 6.6 GW of new and existing clean power by 2035.
- Amazon: invested in X-energy’s small modular reactors (SMRs); the construction permit for its Long Mott site is still under review by the US Nuclear Regulatory Commission (NRC).
Restarting existing plants is fastest and SMRs are slowest: most U.S. SMR power plans target the 2030s. In the near term, AI data centers mainly rely on existing nuclear, natural gas, renewable power purchases, and battery storage. For how SMR technologies differ, see What Is an SMR.
The Power Architecture Is Changing Too, and Taiwan Has a Part
Power isn’t only about “enough or not” but also “how you deliver it.” As rack power moves from a hundred-plus kilowatts toward megawatt class, the traditional power architecture struggles with efficiency and copper losses. So NVIDIA is pushing an 800-volt DC (800 VDC) power architecture that supports racks from 100 kilowatts to over a megawatt, planned for full volume production with the Kyber rack generation in 2027. In August 2026, TrendForce noted that Kyber may slip.
Taiwanese firms have a role in power supplies and cabling: NVIDIA’s published 800 VDC ecosystem list includes Delta, Lite-On, and Richtek (being listed means part of a partner ecosystem, not shipped orders or any investment guarantee). This only describes public supply-chain roles; it does not list beneficiary stocks or give buy/sell advice.
Penchan’s summary
AI’s ceiling is moving from “the chip” to “power.” GPUs still matter, but connecting power, keeping it stable, and getting transformers and interconnection permits are becoming an equally hard, or even harder, limit.
When you read AI data center news, also check where the power comes from and what year it can be connected.
To see how this power is finally carried away as heat, check out What Is Liquid Cooling and What Is Blackwell; to see how all eight gates of the chain connect, head back to the supply-chain overview.