Why is AI almost always sold as a subscription?

Ordinary software costs nothing to run one more time

A spreadsheet or a photo editor is written once and then copied. The hundredth user costs the maker almost nothing, and neither does the hundredth hour of use. That is why software could be sold as a product at a price paid once. Nearly the whole cost sat in building it, not in running it.

A model answering a question is not a copy of anything. It is a computation that has to be performed again, on real hardware, every single time you ask. The cost of later answers depends on the model, provider, caching, and service design. That one fact, more than any business fashion, decides the shape of the price.

What actually costs money when a model replies

A language model writes one token at a time, a token being roughly a short piece of a word. Each token means another pass through the model's parameters, and those parameters have to be read out of high-bandwidth memory for every pass. Each request can use compute, but cost varies with prompt, response length, caching, and service design.

Your conversation also takes up room while it lasts. The model holds a cache of the context it has already read, that cache grows as the conversation grows, and it sits in the accelerator's memory until you are done. Longer contexts can consume more memory and affect capacity, though the impact varies with batching and system design.

• Output length is the main driver of cost: one pass through the network for every token written.
• Long prompts and long conversations cost memory as well as arithmetic, and memory is what limits how many people one accelerator can serve at the same time.
• Sparse designs activate only part of the network for each token, which lowers the cost of an answer without taking it to zero.
• Peak demand sets the size of the fleet, so hardware bought for the busiest hour is paid for during the quiet ones too.

Why a flat monthly fee instead of a price per message

Some providers charge API customers by token, while consumer plans may bundle usage into a recurring fee. Limits, included usage, and overage rules differ, so a flat fee does not reveal the exact cost model.

Some subscriptions use rate limits or usage caps to manage demand and cost. Their presence and effect vary by service; pricing choices also reflect support, storage, product strategy, and operating expenses.

Recurring revenue is genuinely easier to run a company on, and that matters too. But it is not the reason the meter exists. A company using remote inference still pays operating costs, but it can recover them through several pricing models rather than a per-answer charge.

What changes when the computation happens on hardware you already own

Local inference removes the provider’s compute cost for each answer, which can make a one-time purchase practical. It does not determine the license or price. Remote inference has ongoing costs, which a provider can recover through several business models.

Local inference can shift ongoing compute costs from a provider to the device owner, while development, support and distribution still cost money. A provider can choose different licenses and prices; the ownership page discusses those choices.

So can AI ever be bought once?

It depends on the product and its business model. Remote inference has ongoing operating costs, but they may be covered by subscriptions, usage charges, bundles, or other revenue. Local inference removes a provider’s compute cost for each answer, but a seller can still choose a subscription, license, or other price.

Architecture changes the operating costs; it does not determine the license or price. Ask where the answer is produced and how the provider chooses to cover its costs.

When computation happens on a personal device, each answer does not occupy a remote server. That changes the cost structure, though product design, support and distribution still carry costs. The broader point is that where technology does its work shapes the relationship people have with it.