AI Chips: Everything About the Chips That Run AI

Chip brands

Each card uses the same fields so you can compare at a glance. New brands are added as their research is finished and checked.

Start here: how AI chips work

Plain-English explainers that every brand page builds on.

Compare the brands

Headline figures as each company states them. Compare dense with dense and FP8 with FP8: see the glossary below. Rows are added as each brand guide is published.

The same 16 sections for every brand

Learn the layout once and you can read any brand page. Every section has Simple, Deeper and Expert levels, one below the other.

1. 1. At a glance
2. 2. Company card
3. 3. The story
4. 4. The lineup
5. 5. How it’s made
6. 6. Where it’s made: factories and addresses
7. 7. How it works
8. 8. Specs explained
9. 9. Official pricing
10. 10. The software lock-in
11. 11. Who buys it and why
12. 12. Hidden in plain sight
13. 13. Weak spots
14. 14. India angle
15. 15. What’s next
16. 16. Sources

Chip words in plain English

GPU

Graphics processing unit. A chip with thousands of small cores that do many calculations at once, which is what AI training and running need.

Accelerator

Any chip built to speed up one kind of work, such as AI. GPUs, TPUs and LPUs are all AI accelerators.

Fabless

A chip company that designs chips but pays a foundry to manufacture them. NVIDIA, AMD and Apple are fabless.

Foundry

A company that manufactures chips for others, such as TSMC or Samsung Foundry.

Wafer

A thin disc of silicon on which hundreds of chips are made at once, before being cut apart.

Process node

The manufacturing generation used to make a chip, named like "4NP" or "3nm". Smaller usually means more transistors in the same space.

Die

One chip cut from a wafer. Big AI chips now join two dies in one package.

HBM

High-bandwidth memory. Memory chips stacked on top of each other and placed right beside the AI chip so data moves very fast.

Advanced packaging (CoWoS)

The step that joins the chip dies and HBM stacks on one base. CoWoS is TSMC’s version, and it is one of the tightest bottlenecks in AI supply.

FLOPS / petaflops

Floating-point operations per second: how many sums a chip does each second. A petaflop is a thousand trillion per second.

FP4 / FP8 / FP16

How many bits each number uses. Fewer bits means less precision but much more speed. Chip makers quote their biggest numbers at FP4.

Sparse vs dense

A sparse figure assumes half the numbers in a model are zero and can be skipped, so it is often about double the dense figure. Compare dense with dense.

Memory bandwidth

How fast data moves between memory and the chip, in terabytes per second. For running large models this is often the real limit.

NVLink / scale-up

A very fast link joining GPUs inside a server or rack so they act like one big GPU. NVLink is NVIDIA’s; other makers have their own.

TDP / power

The most power a chip is designed to draw, in watts. A whole AI rack can need well over 100 kilowatts.

Rack

A cabinet of servers. Modern AI systems are sold as whole racks, such as a rack of 72 GPUs.