The headline number is always revenue. The number that decides whether it's a good business is what it keeps after the cost of serving customers. For AI companies that gap can be huge, because every answer someone gets costs the company money to generate.
If you came from the video, that's the idea. Here's the part that doesn't fit in sixty seconds: what "cost to serve" actually includes for an AI app, a worked example run end to end so you can see the math bite, and the questions I ask before I believe a margin number at all.
Two words that carry the whole thing
Almost everything you need lives in two terms. Get these and you can read most of what a company tells you.
Cost to serve
What it costs to deliver the product to one paying customer for one month. For an AI app the biggest piece is inference: the bill it pays the model provider every time someone hits generate. It also includes storage, bandwidth, and payment fees that scale up with each user. It does not include salaries or marketing; those come later.
Gross margin
Revenue minus cost to serve, as a percentage of revenue. It's the share of each dollar the company still has before it pays anyone or advertises anything. Seventy percent and up is software-like and healthy. Thirty to sixty is real costs worth watching, where a lot of AI apps sit. Below zero means it loses money on every customer, and growth makes the hole deeper.
Why AI breaks the old rule
For years the rule was simple: software costs almost nothing to serve one more user. You write the code once, and the ten-thousandth customer costs about the same as the first. That's why classic software companies run gross margins in the high eighties.
AI doesn't work that way. Every response burns compute the company pays for in real money, and the heaviest users burn the most. So the cost line grows alongside revenue instead of flattening out. A company can charge a flat monthly price and still have customers who cost more than they pay. That's the reason margin, not revenue, is the number I look at first for anything AI.
A worked example: a subscription logo maker
Abstract math is easy to nod along to and hard to feel, so let's run one. Picture an AI logo-maker app that charges a flat monthly fee and lets you generate as many logos as you want. Every number below is made up to show the shape of the problem, not a real figure from any company.
Say the plan is $19 a month, which looks like a clean software business. Now add the part the pricing page hides: every logo the model draws costs the app roughly $0.12 in image-generation compute. Watch what that does across three kinds of customer.
- The light user makes 20 logos a month. Compute is about $2.40; add card fees and storage, call it $3.50 all in. Gross margin is around 82 percent. Beautiful.
- The average user makes 80 logos. Compute is about $9.60, plus $1 in other costs, so roughly $10.60 to serve. Margin drops to about 44 percent. Fine, but not the software dream.
- The power user makes 400 logos, regenerating endlessly to get it right. Compute alone is about $48; they pay $19. The app loses around $30 on that customer every month, and the more they love it, the more it bleeds.
The trap is that power users are also the loudest fans and the fastest to refer friends. Grow the wrong slice and revenue climbs while the company loses money faster. This is why so many AI apps quietly moved from "unlimited" to credits or usage caps: they had to put a floor under the cost of one customer. My guide on how to choose which AI tool to pay for looks at the same math from the buyer's side.
The worksheet I fill in
Here's the sheet I use on any company I'm curious about. Estimates are fine, just label them so you remember which numbers are guesses. Run it on the logo maker first, then on something real.
Company: [name] Plan I'm modeling: [name of tier]
What one customer pays per month: $[price] [source + date]
Cost to serve that ONE customer per month:
- AI / compute (per action x how many actions): $[..] [estimate?]
- payment + processing fees: $[..]
- storage / bandwidth / other per-user: $[..]
Total cost to serve: $[..]
Gross margin per customer
= (price - cost to serve) / price
= ($[..] - $[..]) / $[..] = [..]%
Now stress it (this is the part people skip):
- Redo the math for a HEAVY user. Still positive?
- What breaks first if the model price doubles?
- What breaks first if compute per action drops by half?
Read it:
70%+ : healthy, software-like
30-60% : real costs, common for AI apps
under 0% : loses money per customer, needs a pricing or cost fixThe line that matters most is the heavy-user check. A margin that only works on the average customer isn't a margin, it's an average hiding a customer who is quietly costing you money.
What the margin number won't tell you
Gross margin is the first number, not the last one, and it's easy to over-trust. A few things I've learned to check before I believe one.
Watch for costs parked below the line. A company can post a lovely gross margin by treating its model bill as a research expense instead of a cost of serving customers. If the compute that answers customer requests isn't in cost to serve, the margin is decorated, not real. Ask what's actually in that line.
Watch for a price propped up by someone else's discount. Early AI apps often ride special model pricing or free credits from a provider. The margin looks great until the promo ends and the real rate lands. When a number seems too good for an AI product, ask what it becomes at list price.
And remember what's still downstream. Salaries, marketing, and research all come out of what gross margin leaves behind, so a thin margin can still lose money after all of it. But if the gross margin is negative, nothing downstream saves it. That's why it's the first thing worth understanding and the last thing to take on faith.
Want to run this on a real company end to end? The research a business guide walks the whole thing through on one company, and what I'd want to know before investing in an AI startup puts margin next to the other questions I ask before I put money in.