ASIC

Every computer contains small components made of silicon that do the computing. They’re called chips. Most of them are generalists: they can do all sorts of things, because software tells them what to do at any given moment. An ASIC is the opposite of that. Here, the task is built directly into the chip’s wiring and can’t be changed afterward. The abbreviation stands for Application-Specific Integrated Circuit, meaning a circuit for exactly one application. A comparison: a Swiss Army knife can do many things reasonably well, a kitchen knife only one thing – but that one thing much better.

Why special-purpose chips are conquering AI data centers

Large AI models essentially consist of enormous amounts of simple calculations, mainly multiplications and additions. A general-purpose processor wastes a lot of energy on tasks nobody currently needs: it has to fetch, sort, and manage instructions. An ASIC dispenses with all of that. Depending on the application, the result isn’t measured in percentages but in multiples: some special-purpose chips perform their task ten times faster or with a fraction of the power.

Electricity has meanwhile become the decisive cost factor. A large data center consumes as much energy as a small town. Anyone who halves consumption per computing step saves nine-figure sums over the years. That’s why Google, Amazon, and Microsoft develop their own chips instead of just buying from Nvidia. Google calls its variant the TPU, Tensor Processing Unit.

For investors, this is a recurring topic in business news. Every report about a tech company’s new in-house chip is also a story about dependency. Whoever designs their own chips negotiates differently with suppliers.

From design to finished circuit

An ASIC is created on a computer. Engineers use a special language to describe which computing units the chip should contain and how they’re connected. Software translates this description into a blueprint made of billions of tiny switches, so-called transistors. This blueprint travels to a chip factory, usually TSMC in Taiwan or Samsung in South Korea. There it’s exposed layer by layer onto silicon wafers.

This process is expensive and slow. The photomasks alone for a modern chip cost tens of millions. From the first design to the deliverable product, it often takes two to three years. And a flaw in the design can’t be fixed with an update – the chip has to be manufactured anew.

That’s why an ASIC only pays off with large quantities or a very high benefit per unit. For smaller projects, there’s a middle path: the FPGA, a chip whose wiring can be reprogrammed afterward. It’s slower and more power-hungry than an ASIC, but flexible. A common misconception, by the way, is mistaking a graphics card for an ASIC. It is specialized for certain types of computation, but remains freely programmable.

Where special-purpose chips are found

ASICs became best known through Bitcoin mining. Starting around 2013, chips that could handle only the Bitcoin computing algorithm pushed all graphics cards out of this business. Such devices can’t do anything else – if the currency becomes worthless, they’re e-waste.

In everyday life, you encounter them without noticing. Inside a smartphone sits a dedicated component for photos, one for video compression, and one for speech recognition. That’s the only reason the battery lasts while filming. Routers, TVs, cars, and hearing aids also contain such circuits.

In market news, the term usually appears in connection with AI accelerators. Companies like Broadcom and Marvell make money designing custom chips for Google or Meta. Anyone reading such news should keep the core point in mind: specialization brings speed and saves power, but costs flexibility and a lot of money upfront.