Peak-Hour Throttling

An online service is not used equally around the clock. There are hours when very many people are making requests at the same time, and hours when almost no one is active. Peak-hour throttling means: during heavily used hours, the provider deliberately slows the service down. This can mean that responses arrive more slowly, that you’re allowed fewer requests per hour, or that a particularly resource-intensive feature is temporarily blocked. The term comes from English: “peak hour” is the busiest hour, and “to throttle” means to restrict or slow down.

Why providers voluntarily slow things down

Behind every AI service are data centers with specialized chips. These chips are expensive and cannot simply be multiplied at will. A provider must therefore decide what load level to build for. If they build for peak demand, the machines sit largely unused for the rest of the day. If they build more conservatively, they save money but have too little capacity during peak hours.

Without a countermeasure, an overloaded service wouldn’t just get slower — it would become unusable for everyone. Requests pile up in a queue, wait times grow, and eventually the connection fails. A planned throttling is the lesser evil. Everyone gets a bit less, but no one is left facing a complete error message.

Then there’s electricity. A data center running at full load draws enormous amounts of energy, and power is often more expensive at peak times. Some operators therefore also throttle for cost reasons, or because the local power grid is reaching its limits. The comparison to a train works well: a train at eight in the morning is packed, at eleven it’s half-empty. Anyone who can shift their trip relieves the burden on everyone else.

Which levers are used

The simplest lever is a counter per user. The service keeps track of how many requests someone has made within a time window. Once the limit is reached, a message appears like “too many requests, please try again later.” Such fixed ceilings are called rate limits. During peak hours, providers set this limit lower than at night.

A second lever is switching the model. Instead of the large, expensive AI model, a smaller one answers the request. This is faster and uses less computing power. The answer, however, is often shorter or less accurate. Many users only notice that the service feels “dumber” on certain days.

A third lever is the order of processing. Paying customers with a contract for guaranteed capacity are served first, while free users wait. Technically, this is a priority queue. It’s important to distinguish this from a network outage: with throttling, everything still works, it’s just deliberately limited. An outage is unplanned; throttling is a decision.

How you notice throttling in everyday life

It shows up most clearly in free versions of chatbots and image generators. There you’ll see messages like “You’ve reached your limit” or “This feature is currently under heavy load.” This usually happens in the late afternoon and evening European time, when Europe and the US are online at the same time. At night, the same services are often noticeably faster.

The principle is also widespread outside of AI. Mobile phone contracts throttle speed after a certain amount of data has been used. Streaming services lower video quality in the evening. Electricity rates become more expensive when many people are charging devices or heating at the same time. Peak-hour throttling is therefore not an AI-specific trick, but an old answer to an old problem.

In business news, the term serves as a warning sign. If a provider throttles frequently, it points to a shortage of chips, data centers, or electricity. That’s exactly why financial media report on it: whoever constantly has to slow down is selling less than they could.