Custom Chip
A custom chip is a computing component that a company has designed for a single, precisely known task instead of buying an off-the-shelf chip. Because it doesn't need to do anything else, it works faster and more efficiently at that task — but development is expensive and time-consuming.
Every device that computes contains small components made of silicon called chips. Most of these are generalists: they are built to handle a wide variety of different tasks. A custom chip takes the opposite approach. It is designed by a company for exactly one purpose and can do almost nothing outside of that purpose. A useful comparison is a kitchen appliance: a multi-function device can stir, chop, and grind, while a pure coffee grinder can only grind — but it does so better, faster, and with less power. This is exactly the principle companies like Google, Amazon, or Apple are selling when they have their own chips developed.
Why corporations invest in their own silicon
The motivation is, first and foremost, simply money. Anyone running an AI model for millions of users has it computing around the clock. At these volumes, every watt saved counts. A chip that uses thirty percent less power per response saves amounts in the millions over the course of a year. On top of that comes data center space: more efficient chips generate less waste heat, and cooling is one of the biggest cost blocks of all.
The second reason is independence. The market for AI computing chips is dominated by a handful of suppliers, above all Nvidia. Anyone buying from them pays high prices and often waits months for delivery. A corporation with its own chip negotiates from a stronger position. It doesn’t need to switch over completely — the mere credible possibility of doing so already drives prices down.
Third, custom chips can be tailored exactly to a company’s own software. A company knows precisely which computing operations its models perform most often. These operations can be built directly into the hardware. A general-purpose chip, by contrast, must be prepared for every conceivable case, wasting chip area and energy in the process.
From design to finished silicon
Almost none of these companies manufacture their chips themselves. They only design the blueprint, often with the help of specialized design houses such as Broadcom or Marvell. The actual manufacturing is handled by a chip factory, known in the industry as a foundry. By far the most important of these contract manufacturers is TSMC in Taiwan. This model is called fabless: designing without owning a factory of one’s own.
The journey from design to the first working sample typically takes two to three years and easily costs several hundred million dollars. A large portion of that goes into the masks, i.e., the exposure templates used in manufacturing. These costs are incurred once, regardless of whether one produces a thousand or ten million chips. That’s why it’s only worthwhile at very large volumes or with very high operating costs.
A common misconception is that hardware alone is decisive. In reality, many chip projects fail because of software. Nvidia has spent years building up CUDA, a programming environment that practically every AI developer knows how to use. Anyone introducing their own chip has to rebuild these tools. Without them, even the best chip remains unused.
Where these chips are already at work
The best-known example is Google’s TPU, short for Tensor Processing Unit. Google has been developing it since 2015 and uses it to train most of its own models. Amazon has two in-house lines, Trainium and Inferentia, one for training and the other for running models in production. Meta, Microsoft, and OpenAI are also working on their own designs.
In everyday life, you encounter this principle in your smartphone. Apple’s M and A chips are custom chips, and they contain dedicated sections just for AI tasks such as facial recognition or speech processing. That’s why on-device photo search works without any data traveling to a data center.
In business news, the term usually comes up in connection with Nvidia’s market position. When a corporation announces its own chip, investors read it as an attack on that dominance. So far, however, custom chips have tended to complement Nvidia rather than replace it: they take over the same recurring standard tasks, while flexible general-purpose chips remain in demand for new research.