AI Startups Risk Becoming Victims of Pricing Power
Set Trending Topics as a preferred source on Google.
Plenty of AI startups think they are running a software business, while their margins read like those of a reseller.
Picture a fitness app that is free to download. Premium costs 9.99 euros a month and includes an AI coach that builds training plans, analyzes runs and answers questions around the clock. 100,000 people use the app and 5 percent pay. That makes 5,000 premium users and nearly 50,000 euros in monthly revenue. It sounds solid, until the bills arrive.
What the Math Looks Like
About 15 percent goes to the app store, roughly 7,500 euros. The AI coach eats an average of 3 euros in token costs per premium user, which adds up to 15,000 euros. Free users would never convert without a taste of the AI features, so they get a few answers per week too, say at 20 cents per head. For 95,000 free users, that is another 19,000 euros. What remains is around 8,500 euros, or about 17 percent of revenue. Premium users pay for their own coach, for the free crowd’s AI answers and for the supplier’s margin. The real winner sits in San Francisco and sends an invoice every month.
The numbers are invented, the mechanics are real. Classic software had marginal costs close to zero, and every additional user cost next to nothing. With AI, every request burns compute. According to a market study, inference accounts for 23 percent of revenue at scaling AI companies on average, and gross margins sit at about 52 percent, compared with 78 to 80 percent for classic SaaS (see the study on the token tax in the application market). That is a structural break in the business model: software that is priced like a commodity trade. Providers such as Notion, GitHub Copilot and Zendesk are already partly moving from flat rates to usage-based pricing, according to Trending Topics.
Who Sets the Prices
Anyone who resells tokens negotiates with suppliers who are also competitors. OpenAI and Anthropic build their own apps, coding tools, agents and browsers. They decide what a token costs, how long a context window can be, which models get retired and which rate limits apply. No startup can tell its customers: we will simply switch suppliers, even though the frontier model of one provider can only be matched by the product of another.
That is pricing power in its purest form. Whoever has the best models can charge premiums, rework price plans or target heavy users. Customers with large bills feel it first. The countermeasure is expensive: the coding editor Cursor is a much-cited example of a startup that tried to break free by investing hundreds of millions in its own model infrastructure. The analysis behind that example concludes that most B2B startups cannot replicate this playbook. Cursor has since become part of SpaceX, which ties the editor to one of the largest AI compute fleets in the world: the group runs its own cluster called Colossus with about 200,000 Nvidia GPUs. Independence from the model providers comes here only as a package deal with a group that brings its own compute.
Few Can Afford Sovereign AI
Real independence means mastering three things: chips, operations and power. That takes GPUs in volumes that create queues outside the factories, data centers with the cooling and grid connections to match, teams that keep clusters running, and energy contracts that stand next to those of heavy industry. Add training, data pipelines and the nerve to burn billions before a single euro flows back.
A handful of corporations and a few extremely well-funded labs can do that. Everyone else, from fintechs to fitness apps, remains a customer. A startup in Vienna, Berlin or Paris training its own frontier model is a lovely thought, but with seed or Series A budgets it is about as realistic as building your own mobile network.
Neoclouds Do Not Solve the Problem
The European answer is often neoclouds: providers that operate GPU capacity in Europe, keep data in the region and sell sovereignty as a feature. That has value for data protection, compliance and latency. Gartner expects such providers to capture roughly a fifth of the AI cloud market by 2030.
For the dependency itself, little changes. Even at a European neocloud, the startup buys tokens or GPU hours and resells them as features to its own users. It is the same middleman model with a different supplier and a different data center location. The margin logic stays, pricing remains with the upstream supplier, and that supplier often relies on US hardware as well. Data sovereignty and economic sovereignty are two different things.
The Counterargument
There is another reading, and it comes with numbers. According to the same market study, inference costs are falling by roughly a factor of ten per year, so gross margins of AI applications should rise. AI-native startups have captured about 63 percent of the application market, and between 71 and 91 percent in areas such as coding, sales and finance. The analysis also notes that established groups keep market power by bundling their products. From this angle, reselling is a transitional phase in which falling token prices and competition among models play into the hands of application-layer startups.
What Founders Can Do About It
Anyone reselling tokens should ask honestly what is left of the product if the supplier offers the same thing tomorrow. Proprietary data, deep workflow integration, trust and distribution are lines of defense that no model update wipes out. Pricing models that reflect usage instead of giving away flat rates keep premium users from financing both the free crowd and the token supplier. Small, cheap models for routine jobs and frontier models only where they make the difference bring the bill down. And architectures that allow switching providers keep the negotiating position open.
A startup that wants to survive stays independent of any single provider, optimizes token costs relentlessly and offers far more value than the raw AI features. Otherwise the whole product can be rebuilt as a ChatGPT plugin in an afternoon.
