Workload

A workload is the work a computer must perform for a particular task. In German one might say Arbeitslast, but in the tech industry the English word has become standard. Both meanings are intended at once: the task itself and the effort it causes. Converting a video is a workload. Running a database for an online shop is another. And training a large language program that can write text is a particularly heavy workload. Whoever talks about workloads is therefore not thinking of tasks as programs, but as portions of computing work that need to be housed somewhere.

Why data centers think in workloads

A data center is a hall full of computers that run around the clock. These computers are expensive, and they consume a lot of electricity. That’s why nobody wants them running idle. Whoever operates the facility must therefore know how much work is currently arising and where it belongs. This is exactly why a term is needed that makes tasks comparable.

Workloads differ greatly in their requirements. Some need mainly computing power, others mainly storage, and still others a fast internet connection. A workload that is allowed to respond a second later is cheaper to run than one that must react instantly. Whoever confuses these ends up paying for technology they don’t actually need.

The word therefore appears constantly in the business reports of major tech companies. When Amazon, Microsoft, or Google talk about AI workloads growing, they mean: customers are running more and more computing work of this kind on their systems. For these companies this is a direct revenue driver, because they bill per unit of computing work performed.

How work is distributed across machines

Between the workload and the actual hardware there usually sits a management layer. It accepts tasks and looks for a machine that currently has capacity. This process is called scheduling. It works similarly to room allocation in a hotel: the guest says what they need, the front desk knows what’s available.

For this to work, workloads are packaged into self-contained units. A common format for this is called a container: a program along with everything it needs to run, in a portable shell. Such a package runs the same way on any suitable machine. If you move it, the program doesn’t notice a thing.

With AI there’s an additional twist. Training a model is a workload that runs for weeks and keeps thousands of specialized chips busy at the same time. Using the finished model is a completely different workload: many small requests, each lasting only seconds. Both require different hardware and different scheduling, even though the same model is behind them.

The word in news reports and invoices

The term appears most often in reports about cloud providers. Cloud means renting computing power instead of buying your own computers. Sentences like “the company is moving its workloads to the cloud” simply mean: the work will in future run on someone else’s machines. Conversely, there are companies bringing workloads back in-house because rental costs have risen.

The word also plays a role in the debate about electricity and climate. AI workloads are considered particularly energy-hungry, which is why new data centers are often built next to power plants. When the news reports that a corporation is securing electricity for upcoming workloads, this is exactly the calculation being referred to.

A common misconception is equating workload with software. Software is the program code; the workload is the running work of that code under real conditions. The same program can be a tiny workload with ten users and a massive workload with ten million users.