Guide
What powers AI: from chips to power plants
Every time you ask a chatbot a question, a chain of very physical things springs into action: chips, memory, buildings full of servers, factories on the other side of the world and power stations. Here is that chain, one layer at a time, with who makes what and why it matters to you.
Your phone or laptop
What is an "AI PC" or an AI phone?
New phones and laptops include an NPU, a part of the chip designed for AI. It runs smaller AI features on the device itself, such as photo editing, live captions and transcription, which is faster and keeps your data private. Big chatbots like ChatGPT still run in data centres, not on your device.
- Microsoft's "Copilot+ PC" label requires an NPU that can do more than 40 trillion operations per second (TOPS). blogs.microsoft.com
Main players: Apple, Qualcomm, Intel, AMD, MediaTek
Why it matters to you: You don't need an "AI" device to use chatbots. When buying, enough memory (16 GB for a laptop) matters more than the AI label.
The data centre
Where does my question actually go?
When you ask a chatbot something, your question travels over the internet to a data centre: a huge building packed with servers full of AI chips, plus the power and cooling to keep them running. The AI model runs there, writes the answer and sends it back, usually in a few seconds.
- Data centres used about 1.5% of the world's electricity in 2024. The US accounted for 45% of that, China 25% and Europe 15%. iea.org
Main players: Microsoft, Google, Amazon, Meta, Oracle
Why it matters to you: Every AI answer costs real money in chips and electricity. That is why the best models sit behind paid plans, and why free plans have limits.
The AI chips
Why do AI companies fight over GPUs?
AI models are built from billions of simple multiplications that can be done at the same time. GPUs, first made for video-game graphics, are designed to do thousands of calculations in parallel, which makes them ideal for training and running AI. Big tech companies also design their own AI chips to cut costs and rely less on one supplier.
- Demand for its AI chips made Nvidia the first company in the world worth $5 trillion, in October 2025. dawn.com
- Google runs much of its AI on its own chips, called TPUs, and rents them to other companies. cloud.google.com
- Amazon designs its own AI training chips, called Trainium. aws.amazon.com
Main players: Nvidia, AMD, Google (TPU), Amazon (Trainium), Microsoft (Maia)
Why it matters to you: Chip supply decides how fast AI improves and how much it costs, which is why chip news moves stock markets.
Memory
Why is memory a bottleneck?
An AI chip is only as fast as the data it can be fed. Large models need to be held in memory right next to the chip, so AI chips use special stacked memory called HBM (high-bandwidth memory). Only a few companies can make it, and making it uses up factory capacity that would otherwise go to ordinary memory.
- Demand from AI data centres is outpacing memory supply, pushing up prices for phones and laptops. Analysts at IDC warn the shortage could last into 2027. 91mobiles.com
Main players: SK hynix, Samsung, Micron
Why it matters to you: The AI boom is one reason phones and laptops have become more expensive, or come with less memory, since late 2025.
The chip factories
Who actually makes the chips?
Nvidia and most other chip companies design chips but don't manufacture them. The most advanced chips are made by a handful of factories, above all TSMC in Taiwan. Those factories depend on machines that print circuits with extreme ultraviolet light, and only one company in the world makes them.
- TSMC makes Nvidia's AI chips, and has started producing them in Arizona as well as Taiwan. blogs.nvidia.com
- ASML, a Dutch company, is the only maker of the extreme ultraviolet (EUV) machines needed for the most advanced chips. asml.com
Main players: TSMC, Samsung, Intel, ASML
Why it matters to you: With production this concentrated, an earthquake, a conflict or new export rules could slow AI everywhere. That is why many governments are paying to build chip factories at home.
Electricity and water
How much power does AI really use?
All those chips need electricity, and they get hot, so data centres also use power and often water for cooling. One AI question uses only a little energy, but billions of questions a day add up, and electricity grids are struggling to keep pace with new data centres.
- Data centres' electricity use is set to more than double to about 945 TWh by 2030, slightly more than Japan uses today, with AI the main driver. iea.org
- Google estimates a typical Gemini text question uses 0.24 watt-hours of energy and about five drops of water. Critics say this leaves out water used at power plants. cloud.google.com
- Microsoft signed a 20-year deal to restart a reactor at Three Mile Island, in the US, to help power its data centres, aiming for 2028. wnn.world-nuclear.org
Main players: Utilities, Nuclear and renewable power companies
Why it matters to you: Your own AI use has a small footprint per question. The bigger question is how countries supply clean power for the total.
What it takes to keep AI running
None of this runs by itself. Construction crews, electricians, chip engineers, researchers and technicians keep the machines built, cooled, repaired and secure around the clock. A few numbers give a sense of the scale:
- $650B+spent in a single yearThe biggest cloud companies are on track to spend more than $650 billion in 2026 on data centres, chips and power, mostly for AI.Source: thenextweb.com
- 122 daysto build one record-breaking sitexAI and Nvidia built the Colossus supercomputer in Memphis, with 100,000 GPUs, in 122 days. Projects like this normally take many months or even years.Source: datacentrenews.uk
- Every 3 hourssomething brokeWhile Meta trained its largest Llama 3 model on 16,384 GPUs for 54 days, there were 419 unexpected interruptions, more than half from faulty chips or their memory. Automated systems handled almost all of them.Source: datacenterdynamics.com
- Up to halfof the cost is peopleTraining the biggest models has become about 2.4 times more expensive every year since 2016. Chips are the biggest cost, but researchers and engineers make up 29–49% of the total.Source: epoch.ai
Do you need an AI laptop or phone?
Probably not yet, if your current device works well. Chatbots such as ChatGPT, Claude and Gemini run in data centres, so they work just as well on an older phone or laptop with a good internet connection. The AI chip (NPU) in newer devices mostly speeds up smaller features such as photo editing, live captions, call transcription and offline translation.
If you are buying anyway:
- Memory matters more than the AI label. Aim for at least 16 GB of memory in a laptop and 8 GB in a phone; AI features, and everything else, run better with more.
- Check which AI features you would actually use and whether they need a paid subscription as well as the new device.
- Battery life, screen and support for software updates still matter more for most people than an AI badge.
How we keep this page current
Each fact links to its source, and we review the page every quarter, because chip and energy numbers change quickly. For what AI can do with all this hardware, see the AI ladder, and for the bigger picture, where AI stands in 2026.