Seventeen Providers, Five Ways to Bill AI, and a Credit That Means Three Things

One provider publishes examples of what its AI credits buy. Another sells a subscription with no unit at all, and reserves “additional charges or suspension” for customers who use too much. Both are selling the same thing: AI features to hosting customers.

Buy AI website building at IONOS and a credit comes with published consumption examples: 150 credits covers two to three new pages or about fifty individual changes. Buy it at Hostinger and a credit is a floating quantity: in the builder’s agentic mode it is deducted to two decimal places, according to the computing resources each request consumes. Run an AI agent at Netlify and a credit is a currency conversion: every dollar of model usage becomes 180 credits, at a rate printed in the documentation.

Three companies, one word, three different contracts. We reviewed the AI pricing pages, help articles and terms of 17 providers, twelve hosting and website-platform brands plus five infrastructure platforms for comparison. Five recurring billing mechanisms emerged, often combined within the same product, with one question underneath them all: when usage grows, who pays.

Key facts

• Five recurring mechanisms: published usage meters, dollar-linked credits, included allowances, variable proprietary credits, and unitless bundles. Some products combine more than one.
• The widest gap: in the offers we reviewed, IONOS publishes examples of what credits buy and does not yet sell top-ups, while GoDaddy’s Airo Plus terms define no unit and reserve “additional charges or suspension” for excessive use.
• The same edit, two credit treatments: manual text changes are explicitly credit-free at IONOS, while edits in Hostinger’s Select & edit mode typically consume 0.01 to 0.05 credits.
• Every client: SiteGround gives every client 20,000 AI tokens a month.
• Free with an expiry signal: Hetzner’s inference API is free while experimental, and its docs already say usage data is kept in order “(in the future)” to “bill usage.”

From 10 Credits Ever to 20,000 Tokens a Month

Compiled from provider pricing pages, help articles and terms, read 25 and 26 August 2026. Full sources below.

One Word, Three Contracts

The credit is the most common proprietary unit in this comparison, and it hides the most. At IONOS it works like a wallet with a published usage guide: the page gives examples of what the credits build, manual text edits are explicitly credit-free, and live sites consume nothing. At Hostinger the same word floats: consumption is “calculated up to two decimal places” according to the resources each request needs, and even manual edits in the visual editor typically draw 0.01 to 0.05 credits each.

WordPress.com says the quiet part in writing. Its Studio documentation explains that what a credit buys “depends partly on what our AI providers charge us, which we don’t fully control.” That is a plain description of the uncertainty inside provider-defined credits whose value is partly driven by model and workload costs: a unit influenced by someone else’s price list, without a published conversion rule the customer can use in advance. Netlify handles the same problem the other way: it puts the conversion rate in the documentation, on the record.

Hard Ceilings, Soft Ceilings, and a Clause Instead of a Ceiling

What happens when the allowance runs out separates the market more sharply than any price does. IONOS applies a hard stop within its capped plans: builds pause until the monthly reset. Top-ups are not sold yet, so the only paid route to more credits is an upgrade to a higher plan. DreamHost, SiteGround and Squarespace sell relief on demand. Cloudflare‘s free tier fails with an error at 10,000 Neurons a day, while its paid tier simply keeps metering.

Netlify gives the ceiling to the customer as a control: set an AI credit usage limit and, once reached, new agent runs are blocked and active runs stop. At the other end sits GoDaddy. The Airo Plus terms name no unit and no number; clause 4.3 requires compliance with “any usage restrictions or fair use policies” and states that excessive use “may result in additional charges or suspension.” The threshold, and the amount of any resulting charge, are not disclosed in the terms or pages we reviewed.

Who Carries the Compute Risk

Each mechanism distributes two different risks: the risk of a higher bill, and the risk that an allowance buys less work than expected. A published meter makes the unit price visible, but the customer still carries the workload risk. An included allowance caps automatic spending, though a variable credit can make its practical value uncertain, and the customer may still choose to buy more. A hard stop protects against overage by converting it into interrupted work. And a unitless bundle leaves the compute cost with the provider, until a fair-use clause lets it charge more or stop the service.

Bluehost is the market in miniature: its three products use two different entry paths and two different treatments of AI usage. Its AI Website Builder is “Build-to-Buy,” free to try for three days, paid only at publish, then carrying an allowance of 20 edit credits a month. Its WordPress AI Builder, the former WonderSuite, ships free with hosting plans and is, in Bluehost’s own comparison table, “Not credit-based.” And its AI Store converts the builder trial into the Sell Online commerce plan, from $15.99 a month on a 12- or 24-month term, or $22.99 month-to-month. One company, three products and two different billing logics.

WP Engine takes a different route with Smart Search AI: pricing tailored to the setup, starting at $140 a month for hosting customers and $150 a month per site standalone, described on its page as flat-rate, with no usage unit or numeric ceiling stated. Kinsta sits outside the billing comparison altogether: the AI-adjacent feature included here, Bot Protection, does not generate anything. It controls AI and automated traffic, and every hosting plan includes it at no additional cost. At one host the comparison concerns AI as a product; at another, a shield against AI traffic.

Free Now, Metered Later

Hetzner is the only provider in this comparison whose entire inference service is free on an explicitly temporary, experimental basis. Its OpenAI-compatible inference API is free while it carries experimental status, rate-limited to 4 million input tokens per minute per key, offered “as is.” The documentation states that usage data is retained in order “(in the future)” to “bill usage,” and that customers will be emailed before the free period ends. The meter is not running yet, but it has already been installed.

Namecheap takes the opposite position: AI building is simply part of shared hosting, no unit, no separate fee, with its WordPress AI tools delivered through the pre-installed Extendify plugin. The public guide does not say whether a separate usage ceiling applies or what happens when AI use becomes unusually high; the fair-use clause elsewhere in this table is one answer a provider has already written down.

Three Questions Before Buying

The comparison suggests a short checklist that works across these mechanisms:

• The unit: what is it, and does the seller publish what it buys?
• The limit: what happens when it is reached: a pause, a purchase, an automatic charge, or a fair-use decision?
• The cap: can the customer prevent further spending?

About the Data

The comparison covers seventeen providers: twelve hosting and website-platform brands and five infrastructure platforms. For each we read the public pricing pages, help articles and terms of service in full on 25 and 26 August, and identified the billing mechanisms used by each product from the provider’s own published wording; the five mechanisms are our grouping, not the vendors’. Kinsta is retained as an AI-adjacent exception and is not assigned to a billing mechanism. Hetzner is listed with the hosting brands as a hosting company, although its reviewed offering, the inference API, is a developer-facing cloud service. Prices and plan terms were taken from the providers’ US-facing or global English-language pages, and regional offers may differ. Every figure was checked against the page it appears on, and figures we could not confirm on a provider’s own pages were left out. Namecheap’s details come from its guide dated July 2024.