Compute (Computing Power)

Every computer works by executing a very large number of simple computational steps one after another. Computing power indicates how many such steps it completes per second. You can think of it like the throughput of a factory: not how clever the work is, but how much of it gets done in a given amount of time. A smartphone manages billions of steps per second, while a large data center facility handles many quadrillions. In technical language this is often measured in FLOPS, meaning floating-point operations per second. When business news talks about “compute,” it almost always refers to precisely this available computing power.

Why computing power became the most important currency of the AI industry

Modern AI systems learn from enormous amounts of data. In the process, a model is repeatedly confronted with examples and improved minimally each time. This process is called training and consists of astronomically many computational steps. For a large language model, thousands of specialized chips run in parallel for weeks to accomplish this. This is precisely why computing power is the central cost factor: it determines who can even participate at all.

The past few years have shown that more computing power usually leads to better models. This relationship is known as scaling laws. Whoever multiplies the amount of data and computational steps reliably gets a more capable system as a result. This turns computing power into a kind of raw material that can be converted into capabilities.

That is why the manufacturers of the corresponding chips have become the most valuable companies in the world. And that is why states now treat computing power as a strategic resource. Export bans on high-performance chips are political news, not merely technical ones.

How chips, electricity, and cooling become usable capacity

Classic main processors, i.e. CPUs, tend to work through tasks one after another. AI computations, however, consist of millions of very similar multiplications. Such tasks can be split up and solved simultaneously. This is exactly what graphics chips, or GPUs for short, are built for. They have thousands of small computing units that work in parallel.

A single chip is not enough for large models. So thousands of chips are wired together into a cluster via very fast data connections. This creates a typical bottleneck: often the computing units are left waiting for data to arrive from memory. Computing power on paper and computing power actually utilized are therefore two different things.

Then there is the physical side. A large data center consumes as much electricity as a small city, and the waste heat has to be removed. Today the limiting factor is often not the chip but the grid connection. That is why new facilities preferably arise where electricity is cheap and abundant.

Computing power in products, prices, and headlines

In everyday life, computing power is noticed indirectly. A chatbot responds promptly or stalls. An image generator delivers the result in seconds or in minutes. Free access is often capped in volume, because every response consumes real computing time. Whoever pays for a subscription is essentially buying preferred access to computing power.

In the news, the term usually appears in the form of investment sums. Companies announce data centers worth billions, or cloud providers rent out computing power by the hour. A common mistake here is confusing computing power with storage space. Storage indicates how much material is available. Computing power indicates how fast that material is worked with.

Also important is the distinction between training and ongoing operation. Training consumes enormous computing power once. Operation, meaning each individual user request, costs less, but does so permanently and millions of times over. Over the entire lifespan of a popular product, operation can end up more expensive than training.