What it costs to run an AI-model business

There is no shortage of content telling you what an AI-model business can earn. There is almost none telling you what it costs. That asymmetry is not an accident: revenue screenshots are good marketing and cost spreadsheets are not.

The result is a predictable failure pattern. Someone reads that top AI creators clear five figures a month, builds a model, reaches a gross figure that looks like success, and only then works out what the platform cut, the chat labour, the tooling and the traffic spend leave behind. The gap between gross and take-home in this business is wide enough to change whether it is worth doing at all.

This page is the cost side. It is organised the way you would organise it in a spreadsheet, because that is the only useful way to think about it.

Every AI-model operation, from a one-person side project to an agency running forty accounts, has the same five cost centres. What changes is their relative weight.

Read that table again and notice what is not at the top. Newcomers obsess over generation cost, because it is the part that feels technical and measurable. It is almost always the smallest controllable line.

Fanvue's own documentation describes a 20% commission across subscriptions, tips, one-time payments and paywalled content sales. Secondary write-ups we have seen, none of which we would cite as a source, have variously claimed 15% for a first period and 20% thereafter. We have not been able to reconcile those claims against Fanvue's published creator terms, so treat any figure below 20% as unverified until you see it on your own dashboard.

The practical point: the platform fee is the only cost centre you cannot optimise after the fact. You pick it when you pick the platform, and switching later means rebuilding your subscriber base. Get this decision right before you get anything else right. Our platform take-rate comparison works through the net-to-you arithmetic across the main options.

This is the line most first-time models get wrong by an order of magnitude. On subscription platforms, a large share of revenue comes from paid messaging: pay-per-view unlocks, custom requests and tips generated in conversation. Conversation requires someone to have it.

That gives you three options, and all three cost real money.

You do it yourself

Cash cost of zero, which is why everyone starts here. What it costs instead is that messaging volume grows with subscriber count, so your own time becomes the binding constraint on revenue. Most solo operators hit that wall somewhere in the low thousands per month.

Hire chatters

Typically priced as a share of the revenue they generate, or hourly for shift coverage. This converts your time constraint into a margin constraint, which is usually the right trade. It also introduces a management job you did not previously have, plus a real trust and access-control problem.

Agency management

Someone else runs the whole revenue operation for a share of gross. Simplest to operate and by far the most expensive. Also the option with the widest quality distribution: the agency layer in this niche includes both competent operators and outright bad actors.

Here is the arithmetic that surprises people. Runpod's published Secure Cloud rate for an RTX 4090 is $0.69 per hour, with Community Cloud listings for the same card at roughly half that. A tuned image pipeline on that class of card produces images in seconds, not minutes.

Runpod Secure Cloud, RTX 4090Published rate, August 2026

$0.69/hr

Community Cloud, same cardFloating marketplace rate, host-dependent

~$0.34/hr

H100 PCIe 80GB, Secure CloudRelevant for training, not for inference

$2.89/hr

Even generating several thousand images a month, rented GPU time is usually a two-figure monthly bill. Compute is not the cost that matters. The keeper rate is: the proportion of generated images good enough to publish. A pipeline with a 5% keeper rate costs twenty times more per usable asset than one at 100%, on identical hardware. That is where the real money is won or lost, and it is the subject of our cost-per-usable-image model.

This line is either zero or enormous, with very little in between. Organic distribution across social platforms costs time rather than money. Paid acquisition into adult-adjacent offers is expensive, because the mainstream ad platforms will not take the spend and the ones that will charge accordingly.

The trap is treating paid traffic as a growth lever before you know your conversion rates. Without a measured subscriber value, paid acquisition is not growth, it is a subsidy you are paying to strangers. Establish the organic funnel first; it also tells you whether the offer works at all.

The boring line. It is small, it is mostly fixed, and it is the one people over-buy because subscribing to software feels like progress.

• Generation tooling or a studio subscription
• Scheduling and posting tools, if you are running several channels
• Storage for the asset library, which grows faster than expected
• A password manager and a separate identity stack, which is not optional if anyone else touches the accounts
• Accounting, and eventually an accountant who understands platform income

The last item on that list is the one that decides the others. Platform income arrives net of fees, often across currencies, sometimes with holds. Reconstructing a year of it in April is miserable. Set up the bookkeeping in month one.

Structure matters more than precision here. The categories below are the ones we would insist on seeing before believing anyone's claimed margin.

GROSS REVENUE
  Subscriptions
  Pay-per-view unlocks
  Tips
  Custom content
  Referral / affiliate income
= GROSS

DIRECT COSTS
  Platform commission            (% of gross, per platform terms)
  Payment / payout / FX fees     (easy to miss, non-trivial)
  Chargebacks and refunds        (track separately from fees)
= NET REVENUE

OPERATING COSTS
  Chat labour                    (record hours AND attributable revenue)
  Content generation             (GPU hours + API spend)
  Traffic acquisition            (paid only; log organic hours separately)
  Tooling and subscriptions
  Storage and infrastructure
= OPERATING PROFIT

NON-CASH / DEFERRED
  Unpaid founder hours           (log them; a business that only works
                                  because you are free is not working)
  Tax provision                  (set aside monthly, not at year end)

1. 01

   Cash break-even

   Gross revenue covers platform fees, tooling and any paid labour. This is the threshold everyone tracks, and it is the least informative of the three because it ignores your own time entirely.

2. 02

   Labour break-even

   Revenue covers all cash costs and pays your own hours at a rate you would accept from an employer. An operation can look profitable for a long time without crossing this line.

   Gotcha — Calculate this before you decide to scale. Scaling a business that has not reached labour break-even multiplies the loss.

3. 03

   Delegable break-even

   The operation covers all cash costs, your hours at market rate, and a replacement for your hours. This is the point at which you own an asset, not a job, and it is the only threshold that makes the business sellable.