Revenue Is Not Margin: Four AI-Earning Ledgers After Costs
Anicca booked $230.83 of revenue against $679.88 of expenses in one month, netting minus $449. Krimskrams earned $1.54 in three weeks while burning 5B+ tokens, about $7,000 at API prices. BugBasher sold pest-control leads for $75 over four weeks, against roughly $3,000 and thousands of phone calls. All of them reported revenue. Few reported cost.
📜 Table of Contents
Four ledgers, four ways to lose
Anicca: 1 to 30 October 2026, $230.83 of revenue against $679.88 of expenses, a net loss of $449.05. Of the revenue, $203.86 came from freelance gigs and $26.97 from an agent marketplace.
Krimskrams: pay-per-call data endpoints for other agents, 17 products, $1.54 in three weeks. In the same period it burned 5B+ tokens, about $7,000 at API prices (the team actually paid $400 a month in subscriptions plus $47.19 of its own money).
BugBasher: sold pest-control leads through Devin at $75 each. One sale in four weeks, $75 total. The authors say it cost roughly $3,000 and thousands of phone calls.
autocomp: 39 days of an open-source "company" loop, 487M tokens, about $1,117 in model cost, zero revenue.
Four projects, two accounting styles. Anicca and autocomp published cost, so the net loss is calculable. Krimskrams and BugBasher are retrospective estimates.
Why these projects lose money so easily
Ordinary businesses lose on inventory, rent and labour. AI projects lose somewhere most people never look: inference cost scales linearly with usage, and the meter runs in the background.
Three specific patterns:
Retry amplification. One failed delivery, three retries, four times the cost. A task with a 60% success rate actually costs 1.7x its sticker price.
Context inflation. In multi-turn work, every turn resends the history. By the last turns of a long task, a single call can cost ten times the first.
Idling. Anicca's agent kept trying, failing and retrying with nobody calling stop. Of autocomp's 487M tokens, nobody has broken out how much was useful work.
Four numbers before you start
One: inference cost per delivery. Not monthly bill divided by orders. Per task: average tokens, tool calls and retries for one successful delivery. This sets your price floor.
Two: actual revenue per delivery. Subtract refunds, platform cut and payment fees. Anicca's $230.83 is gross; after platform fees it may be 70% of that.
Three: gross margin. (Revenue - inference cost - platform fees) / revenue. Below 50%, an AI service business struggles to cover acquisition.
Four: orders to break even. Fixed costs (subscriptions, servers, domain) divided by gross margin per order. Krimskrams has $400 a month of fixed cost; at $0.09 margin per order it needs 4,400 orders.
The counter-example
Of the 76 businesses in the Agent Earnings directory with verified revenue, the median is $4,380 a month. What they share is not better technology. It is higher prices.
The same model call behind a $9 tool and a $400 monthly retainer costs about the same to run. The revenue differs by 40x. The difference buys accountability, industry knowledge and trust, none of which are expensive in tokens.
So the survivors almost all run high-price, low-volume, human-backed services rather than low-price, high-volume, fully automated tools.
My read
"The marginal cost of AI approaches zero" is false at the inference layer. Tokens are metered, every call costs real money, and stronger models cost more per call.
What does approach zero is replication cost: a prompt, a template, a workflow costs almost nothing to hand to the second client. But the savings tend to get eaten by acquisition cost.
If you want to run one of these, write the four numbers down first. If you cannot produce them, do not start.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.