Software gets the headlines — hardware does the work
When you type a question into ChatGPT or Claude, the answer doesn't come from your phone or laptop. It's processed on specialised computers in a data centre, often thousands of miles away, built specifically to do the kind of maths AI requires.
Why AI needs different hardware than a normal computer
A regular computer chip (a CPU — Central Processing Unit) is built to do many different tasks, one or a few at a time, very precisely and in strict order. That's great for running an operating system or a spreadsheet.
AI training and inference (the technical term for "using" a trained model) involves something different: an enormous number of simple mathematical calculations — mostly multiplication and addition — happening at the same time, in parallel. A CPU can do this, but slowly. This is where GPUs come in.
A useful analogy: a CPU is like one highly skilled chef cooking one complex dish very well. A GPU is like a thousand line cooks each doing one simple, repetitive task — chopping one vegetable each — all at once. For AI, you usually need the second kind of kitchen.
The three types of chips that matter in AI
- GPU (Graphics Processing Unit). Originally built for rendering video game graphics, GPUs turned out to be extremely good at the parallel maths behind AI. Nvidia is the dominant maker of AI-grade GPUs today.
- TPU (Tensor Processing Unit). A chip Google designed specifically for AI workloads, rather than adapting a graphics chip. "Tensor" refers to the type of mathematical structure used throughout machine learning.
- Custom AI chips. Several large AI companies now design their own chips tailored precisely to their own models — Amazon, Microsoft, and others have their own in-house silicon for this reason.
What a data centre for AI actually looks like
Training a large model isn't done on one chip — it's done across thousands of GPUs or TPUs working together, connected by extremely fast networking, often filling entire warehouses. These facilities need:
- Massive electricity supply — training a very large model can use a huge amount of electricity over the weeks it runs.
- Serious cooling systems — this many chips running at once generates enormous heat, often requiring liquid cooling rather than just fans.
- High-speed interconnects — the chips need to constantly share data with each other during training, so the connections between them matter almost as much as the chips themselves.
Training vs. inference — different hardware needs
- Training — building the model in the first place — is the most hardware-intensive part, often taking weeks across thousands of chips.
- Inference — actually using the trained model to answer your question — needs far less hardware per request, but happens constantly, at huge scale, across millions of users simultaneously.
Why this matters even if you're not technical
Chip availability is one of the biggest real-world bottlenecks in AI right now — it is one reason new AI features sometimes roll out slowly, why heavy AI use can cost money, and why countries and companies compete over chip manufacturing and supply.
Latest: Hardware Announcements
Checking for the latest hardware announcements…
Frequently asked
Can AI run on a normal laptop?
Smaller AI models can run on a capable laptop, especially with a decent GPU. Large models like the ones behind ChatGPT or Claude need data-centre-scale hardware — far beyond any personal computer.
Why is Nvidia mentioned so often in AI news?
Nvidia makes the large majority of the high-end GPUs used for AI training worldwide, which has made it one of the most valuable companies in the world during the current AI boom.
Is the hardware news on this page live?
The "Latest: Hardware Announcements" section above checks automatically about every 15 minutes for new, verified hardware announcements from trusted sources (Nvidia, AMD, Intel, TSMC, major cloud providers, or Reuters). The lesson content itself is written to stay accurate regardless of any single announcement, so it's reviewed manually rather than auto-rewritten.
← Lesson 002 See Current AI Hardware News →