TL;DR
Normal software is cheap to serve: one more customer costs almost nothing. AI is different. Every time a user asks the AI something, the startup pays a bill to the company that runs the AI model. Because of this, AI products keep about 52% of their revenue after direct costs, while normal software keeps 75–85%. AI prices are falling fast, but usage is growing even faster, and the companies that set the prices also sell competing products. Founders can protect themselves by tracking AI costs closely, charging based on usage, and building something a model update can’t copy.
The simple idea
Think of a normal software company like a book publisher. You write the book once, then print copies for a tiny cost. The more you sell, the more profit you keep.
Now think of an AI startup like a restaurant that buys its main ingredient from one supplier. Every plate it serves costs real money. If the supplier raises prices, the restaurant’s profit shrinks the same day.
That’s the situation for many AI startups. They don’t build the AI “brain” themselves. They pay another company, like OpenAI or Anthropic, every time their product uses it. That payment is called an inference cost, and it is charged per “token” (a token is roughly a small piece of a word).
What the numbers say
A January 2026 survey by ICONIQ Capital asked about 300 software executives building AI products. The key results:
- AI products keep about 52% of their revenue after direct costs in 2026. That’s up from 45% in 2025 and 41% in 2024, so it’s improving.
- At companies that are growing fast, about 23% of revenue goes just to running the AI.
- Normal software companies typically keep 75–85%.
Another study, by Bessemer Venture Partners, looked at the fastest-growing AI startups. Many of them kept only about 25% in their early days. The steadier ones kept about 60%.
In plain terms: for every $100 an AI startup earns, it keeps roughly $52 to pay for everything else, compared with $75–85 for a normal software company.
An example (numbers are made up)
Imagine a startup that sells an AI tool to review contracts. It charges $99 per month and has 2,000 customers.
- Money coming in: 2,000 × $99 = $198,000 per month
- Each customer reviews about 120 documents a month, and each review costs the startup about $0.35 in AI fees. That’s 2,000 × 120 × $0.35 = $84,000 (about 42% of revenue)
- Other direct costs like hosting and support: about $26,000
- What’s left: $88,000, or about 44%
Now the AI company raises its prices by 25%. The AI bill goes up to $105,000. What’s left drops to about $67,000, or roughly 34%.
Nothing changed about the startup’s customers or product. One price announcement from a supplier cut its profit by almost a quarter. Fixing it by raising prices or redesigning the product takes months.
“But AI is getting cheaper, right?”
Yes, and very fast. Researchers at Epoch AI found that the price of getting a fixed level of AI quality has dropped between 9 and 900 times per year, depending on the task. Stanford’s AI Index reported that the cost of a GPT-3.5-level answer fell from $20 per million tokens in late 2022 to about $0.07 by late 2024. One index of market prices reportedly hit a record low of about $0.97 per million tokens in September 2026.
So why do startups still feel squeezed? Three reasons:
- Customers want the best AI, not the cheap one. The big price drops are for older, “good enough” models. Users compare your product with the newest and best, which costs more.
- AI now does bigger jobs. New “AI agent” tools take many steps for a single task, using lots more tokens each time. The price per token drops, but the number of tokens per task goes up.
- Flat plans don’t match variable costs. If you charge everyone $20 a month but one user costs you $200 in AI fees, that user is a loss.
Lessons from AI coding tools
AI coding assistants show this problem most clearly, because they use a huge number of tokens.
- Replit’s profit margin reportedly swung between about -14% and 36% during 2025.
- Cursor’s company apologized in 2025 after a pricing change surprised users with extra charges. It then signed deals with OpenAI, Anthropic, Google and xAI for a higher-limit plan.
- Founders told TechCrunch that margins on many AI coding products were close to zero or negative. The AI model companies were also launching their own coding tools, which means their customers’ supplier is now also a competitor.
Costs outside the AI models are rising too. One report said AWS raised prices on its reserved AI chip capacity by about 15% in early 2026.
The hopeful side
It’s not all bad news. Margins have been going up every year. Prices keep falling. One projection suggests that startups mixing open-source and paid models could reach around 85% margins by 2027. And AI startups are winning most of the new application market, not losing it.
The real story is a race: falling prices on one side, and growing usage and rising customer expectations on the other. Whoever manages the second part gets the benefit of the first.
What founders can do
1. Know your AI cost for every customer and feature. Don’t just look at one monthly bill. If you can’t see where the money goes, you can’t price properly.
2. Use the cheapest model that does the job. Simple tasks like sorting or formatting can use a small, cheap model. Save the expensive model for the moments where quality really matters.
3. Charge based on usage. A flat price works only if your heaviest users don’t wipe out your profit. A base fee plus extra credits for heavy use protects you. Be careful with free plans: free AI users are pure cost.
4. Don’t depend on one supplier. Build your product so you can switch between AI providers. Even if you never switch, being able to switch gives you bargaining power.
5. Build something a model update can’t copy. Your own data, deep integration into how customers work, strong distribution and trust are hard for a supplier to copy. If someone could rebuild your product as a plugin in an afternoon, you have a bigger problem than costs.
The bottom line
The old rule of software, “more customers means more profit,” doesn’t work automatically for AI. Every use costs money, and the supplier controls the price. AI startups can still do very well, but only if they treat the AI bill as a core part of their business from day one: measure it, price around it, and build enough of their own value that a supplier’s price change is an annoyance, not a crisis.
Related Buzz: We also covered [Your Amazon Seller Dashboard Now Talks to Claude — Here’s What That Means]