Running the numbers

The target is small. That is the point.

Two million in annual recurring revenue at twenty five dollars a month is six thousand six hundred and sixty seven customers. Set against thirty six million small businesses in the United States, that is eighteen thousandths of one percent.

So market size is not the question worth modelling. The question is whether roughly three hundred new customers a month can be acquired, every month, for two and a half years, at a cost that the contribution margin can carry. This page models that, and separates what is published from what is guessed. Mail and domain registration are deliberately out of scope: the pod needs authoritative DNS, nothing more.

Read the colours. Petrol values come from published rates or federal data. Amber values are assumptions with no source behind them, and every one of them is a thing to go and find out.

Where the customers come from

Three pools, one rounding error.

These are the populations, not the addressable market. The useful thing they show is how small a share of any of them the plan actually needs.

Wave one · the tinkerers
50M+
people already paying monthly for a frontier AI assistant
  • OpenAI confirmed more than 50 million paying consumer subscribers across its tiers in April 2026
  • Anthropic does not publish a consumer count; outside estimates put total Claude consumer users between 18 and 30 million, and the company said paid subscriptions more than doubled during 2026
  • Google reports more than 350 million paid subscriptions across its consumer services, though that figure includes storage plans rather than AI alone
Wave two · small business
36.2M
small businesses in the United States
  • The SBA Office of Advocacy counts 36,207,130 small businesses, of which roughly 29.8 million have no employees at all
  • About 5.62 million new business applications were filed in 2025, running at roughly double the pre pandemic rate
  • Solo ventures are the overwhelming majority, which is exactly the segment that has never been worth a support call
Waves three and four · builders
47.2M
active developers worldwide
  • SlashData put the global active developer population at 47.2 million in early 2026, counting anyone who writes code monthly
  • Evans Data counts only paid professionals and arrives at 35.6 million, growing at roughly 5.5% a year
  • GitHub passed 518 million accounts in 2025, which is the ceiling on people who could plausibly want version history as a product feature
The model

Move the variables. Watch what breaks.

Two paths to the same revenue number. The first buys customers. The second already has them, which is the entire argument for building this inside an established host rather than beside one.

Revenue per pod

The primary dial. Everything below is a rate applied to this.
The site currently shows nineteen dollars as illustrative. The model has an opinion about that.
No source. Developers on consumption pricing could land anywhere.

Cost per pod

1.0 is published pay as you go container rates. An operator on its own hardware pays less, which is the whole point of pitching one.
The largest unknown in the model, and the only cost that rises with the feature people came for.

Buying growth

Seventy five is the working assumption for paid social and creator content. Drag it to 150 and watch the ramp, not the ratio.
Resellers cut acquisition cost by roughly two thirds and take a wholesale margin. The model applies both.
Annual recurring revenue
Gross margin per pod
Contribution per pod
per month, after cost of service
Payback
months to earn back acquisition
Lifetime value to cost
Time to two million in recurring revenue
Where a pod's cost goes

The ramp

Discipline

What is known, and what I made up.

Any operator worth pitching will separate these two lists within about ninety seconds. Better to do it first.

Published or federal data
  • Container compute. Two cents per CPU core per hour and half a cent per gigabyte of memory per hour, from bunny.net's published Magic Containers rates. Memory, not CPU, is what makes an idle pod expensive.
  • Persistent storage. Ten cents per gigabyte per month on provisioned volume size, though that meter is not yet billing.
  • Egress. One cent per gigabyte in Europe and North America, rising to six cents in the Middle East and Africa.
  • Domains, if they are ever added. The .com wholesale registry fee is $10.26 a year, going to $10.97 on 1 November 2026, plus the ICANN fee. Not in the bundle, so not in the cost stack. Kept here because an add on still has to price above cost.
  • Certificates. Zero. This used to be a line item and no longer is.
  • Small business counts. 36,207,130 businesses and 5.62 million new applications in 2025, from the SBA Office of Advocacy and Census Business Formation Statistics.
  • Developer counts. 47.2 million active developers per SlashData, 35.6 million paid professionals per Evans Data.
Assumptions with nothing behind them
  • Inference cost per pod. The default assumes roughly twenty assistant turns a month with prompt caching. A chatty customer could be ten times that. This needs a real measurement from a real prototype before it goes in front of anyone.
  • Acquisition cost. Seventy five dollars is a hope, not a figure. It is testable for a few thousand dollars and should be tested before it is presented.
  • Churn. Two and a half percent monthly is a guess in the right neighbourhood for small business hosting, and the number that most affects the answer.
  • Residual support. The premise is that the assistant absorbs routine questions. Nobody has measured how much.
  • Measured pricing revenue. No idea what a developer on consumption pricing actually bills. Wide enough range to be nearly meaningless.
  • Migration take rate. What share of an existing shared hosting base would accept a move, and how many would use the moment to leave instead.
The conclusion the model keeps reaching

The economics work either way. Only one path can actually be staffed.

At any sensible price above the mid twenties, contribution per pod carries a seventy five dollar acquisition cost several times over, and carries a hundred and fifty dollar one comfortably. The unit economics are not the problem.

The problem is throughput. Three hundred net new customers a month, sustained for two and a half years, is a marketing operation. An established host with a few hundred thousand existing customers reaches the same revenue by moving two tenths of one percent of its base per month, at a fraction of the cost, with the churn risk pointing the other way.

Read the commercial model for how the billing works, and the open questions for what is still unresolved.