AI EncyclopediaFirst steps5/5
First steps in chips and AI infrastructure
A guide for complete beginners
This field covers the chips (semiconductors) that compute and remember, and the data centers and electricity that run AI on those chips.
Why everyone is talking about it
AI runs on chips in the end. Training and using a large AI takes huge numbers of AI chips and fast memory, so who designs the chips and who makes them has become the center of the AI race.
A data center holding tens of thousands of chips can use as much electricity as a small city. That is why chip news now also covers factories and equipment, export controls between countries, and power grids and cooling.
The big picture
- DesignDraw the chip's circuits. A company that only designs and leaves production to others is called fabless.
- ProductionA foundry etches very fine circuits onto a wafer to make the chips. Equipment and yield decide how good it is.
- PackagingFast memory is attached next to the computing chip and the two are packaged as one.
- Data centerTens of thousands of servers with chips installed are gathered in one place to train and run AI.
- Power and coolingAI infrastructure also means cooling the heat the chips give off and supplying electricity that never cuts out.
How one chip ends up running AI, shortened to five steps.
Six ideas to know
Read them in this order and the news gets much easier. Each one opens its encyclopedia entry.
- SemiconductorA material that can either pass electricity or block it. The small chips made from it are also called semiconductors.
- GPU (graphics processing unit)A chip that does the same calculation a huge number of times at once. It is the chip most used to train and run AI.
- HBM (high bandwidth memory)AI memory built by stacking memory chips in layers so that data moves in and out very fast.
- FoundryA factory that makes semiconductors designed by other companies, or that business.
- Process nodeA name for how thin and tightly packed a chip's circuits are made. The smaller the number, the more advanced the technology.
- AI data centerA large facility that gathers tens of thousands to hundreds of thousands of high-performance chips in one place to train and run AI.
Common misunderstandings
Myth"3 nm" means the circuits are really 3 nm wide.
In fact These days it is closer to a label that names a generation. Each company uses different standards, so it is hard to compare by the number alone.
MythWith a good chip alone, you can run AI.
In fact Memory that feeds data to the chip quickly, links between chips, electricity and cooling all have to support it. That is why memory and data center companies get attention too.
MythOne country or one company makes all of a chip.
In fact Design, equipment, materials, production and packaging are spread across many countries and companies. If even one is blocked, the whole chain shakes, so supply chains and export controls are often in the news.
Who to watch now
Follow on the tech radar
News about each technology, company or person, in date order.
Recent milestones
The three latest events from the timeline. The timeline shows how the field got here.
Where to go next
This guide was written by AI and checked again for accuracy before publishing, and the English edition is translated from the Korean guide. Rankings change quickly, so they are not written into the text; 'Who to watch now' shows the LastAI Index as it is. Tell us if something is wrong. Report an errorLast updated