Usage-Based Billing

Usage-based billing means: the price depends on how much someone actually uses an offering. Those who use little pay little. Those who use a lot pay a lot. The counterpart is the flat rate, such as a subscription for 9.99 euros a month, which always costs the same regardless of usage. Everyday life offers a familiar example in electricity or water: the meter measures consumption, and the bill arrives at the end of the year. Providers of computing power and internet-based programs have adopted exactly this model.

Why cloud and AI providers sell this way

For providers, this model solves a cost problem. Every request to an AI model incurs real costs, because somewhere a data center is consuming electricity and running expensive specialized chips. With a flat rate, heavy users would be a loss-making business, since they consume more computing time than their subscription brings in. With usage-based pricing, revenue automatically grows along with costs.

For customers, this lowers the barrier to entry. A startup with three employees doesn’t have to sign a contract worth 50,000 euros a year just to try out an AI. In the first few weeks, it might pay twelve euros. Only once the product is running and many people are using it do costs rise. The risk thus shifts from the customer to the provider.

This also plays a role on the stock market. For a classic subscription provider, next quarter’s revenue can be predicted fairly well, because the contracts are fixed. For usage-based providers, it fluctuates more strongly, because if customers cut back, revenue drops immediately. Analysts therefore pay attention to metrics such as net revenue retention, which shows whether existing customers are expanding or scaling back their usage.

What exactly gets counted

First, a unit of measurement is needed. For language models, these are usually tokens, i.e. word fragments. One token corresponds to roughly four characters of text. Both the input text and the generated response are billed, often at different prices. For storage services, gigabytes per month are counted; for virtual servers, runtime in seconds or hours.

Then the provider has to keep count. Every request is logged and assigned to an account. This counting data flows into a billing system that generates an invoice from it at the end of the month. Because large services generate billions of individual events, this system is technically demanding. There are entire companies that offer nothing other than software for exactly this kind of counting.

In practice, pure models are rare. A hybrid form is common: a base fee plus consumption, or an allowance included in the price. If you use more, every additional unit is billed separately. Many providers also offer volume discounts, so that the millionth request costs less than the first.

Between free allowance and cost trap

You encounter this model most directly when using an application programming interface yourself, i.e. a technical access point through which your own programs can address an AI model. Providers such as OpenAI, Anthropic, or Google publish price lists for this per one million tokens. Cloud platforms such as AWS or Microsoft Azure also bill this way. Even someone who only pays for a chat subscription indirectly encounters it, since many subscriptions include a usage limit.

The best-known drawback is the billing surprise. A faulty program that sends requests in an endless loop can rack up high costs overnight. That’s why reputable providers offer spending limits and warning alerts. A typical misconception is assuming that usage-based is automatically cheaper. With very consistent, high usage, a flat-rate contract is often cheaper.

In business news, the term usually appears in connection with quarterly results. Reports such as “growth is slowing because customers are optimizing their cloud spending” refer to exactly this model. Anyone who understands that revenue here is directly tied to consumption can interpret such news considerably better.