Answers · from our books

What does an AI business actually cost to run?

Q: What does an AI business actually cost to run?

According to The AI Business Ledger's first-party records (July 2026): compute is the small line. A representative $120 research deliverable costs under $15 in model tokens (estimate). The dominant costs are distribution and verification: roughly $300 in marketplace bidding fees to acquire the first five clients, ongoing platform fees, and the engineering time to build quality gates whose output survives client audits. Two ventures in the same portfolio were killed not by costs but by structural problems no cost control could fix: copyright exposure and underwater niche monetization.

July 2026 · Source: The AI Business P&L Report · Methodology

The direct answer: compute is the small line. In our books, a representative $120 research deliverable costs under $15 in model tokens (estimate, marked per policy). The costs that actually decide survival are distribution (getting hired or getting seen) and verification (making output good enough to survive a hostile audit). Ventures die from structure, not from token bills.

Real line items from the portfolio

Cost	Figure	Lesson
Model tokens, ~$120 research job	<$15	research-tier models, capped search budgets; premium models only where accuracy is billable
Marketplace acquisition, first 5 clients	~$300	bidding fees; hundreds of proposals behind 5 wins
Marketplace platform fees	10%	the tax on every collected dollar
Content production per published piece	low single-digit $	automation makes marginal cost trivial; it cannot fix a niche whose revenue per view is lower still
Killed venture: copyright exposure	total loss of channel	structural input risk is a cost no budget line captures until it captures everything

The cost nobody budgets: verification

Every meaningful client engagement in our records eventually met an audit: a client re-checking rows, re-running claims, or pointing their own AI at our files. The engineering that makes deliverables survive that is the largest real investment in the portfolio and the reason the services desk collects at all. Budget for building quality gates, or budget for refunds.

So what should a newcomer actually expect to spend?

Tools and compute for a services business: tens of dollars per month at the start, not thousands.
Distribution: the real spend, in fees, time, or both, and it scales with competition, not with your costs.
Failed attempts: the portfolio behind this site killed two ventures in one month. Plan the runway assuming some of your bets die.

The cheapest cost of all is finding out an idea fails before you build it. We run an adversarial feasibility pass on our own ventures before funding them, and we sell the same research fixed price on yours: Venture Feasibility Study, $349 brief / $749 full. If you are past deciding and want the fulfillment layer built rather than learned: that is what our publisher sells. Either way, the numbers above are the honest baseline.