AI has a getting-paid problem
The industry is racing to cut token costs. But cheaper compute won’t fix failed payments, lost renewals or customers who can’t pay.
This summer, Rippling found one of its engineers was spending $50,000 a month on AI. In fact, its token bill was on pace to consume 40% of its R&D headcount budget. And Rippling was not alone.
Today, every prompt a customer types and every answer a model returns is counted in tokens, and every token has a price. In June, Cloudflare added spend limits to its AI gateway. By August, Databricks, Rippling and Ramp had all launched tools to budget, track or reroute that spend.
The rush makes sense. Inference - the computing work a model does every time it answers a prompt - grows with every customer, and a single heavy user can cost more than they pay. But the rush treats AI's money problem as a cost problem only. For AI companies, margin is set twice: once by what each output costs to produce, and again by whether the customer's payment goes through. That second part gets far less attention, and it is where I see AI companies leave revenue behind.
Global-first by default
Lovable, the Swedish company that lets people build apps by describing them, sells to businesses and consumers anywhere, and anyone with a card can start paying. In fact, recent industry analysis found AI businesses selling into 55 countries in their first year, against 25 for SaaS. Paying users arrive in dozens of markets before anyone has chosen an acquirer, a payment method or a tax registration for them.
That gap is also where payment risk builds. Most young AI companies route every payment through one acquirer at home. To a bank in São Paulo or Munich, each charge looks foreign, and foreign card payments are declined more often.
What makes the matter more interesting is that AI companies by design have less margin to absorb the loss. ICONIQ's January 2026 survey put average gross margin for the AI-native products it tracks at 52%, against 75% to 85% for traditional SaaS. When a payment fails, the compute has usually already run. Which means, the company loses the sale and keeps the cost.
Three places where AI companies leak revenue
Especially when their payments are not built for global reach.
- Failed payments. Cross-border routing, expired cards and weak retry logic turn good customers into declines.
- Missing local payment methods. India was already ChatGPT's second-largest market, yet 29 million downloads in 90 days had earned just $3.6 million. OpenAI's answer, ChatGPT Go, cost ₹399 a month and accepted UPI.
- Tax and compliance friction. Selling digital services abroad can trigger VAT or GST obligations in more than 110 jurisdictions. For a company that launched globally, those obligations arrive long before any local team does.
How an AI company should think about its payments
Most people file AI companies as subscription businesses. The reality is that users often run out of credits mid-task and buy more. Speaking at a payments conference this year, Lovable's head of growth, Elena Verna, said its "repurchase rate of top-ups was just as high, if not higher than subscription renewal rates."
That gives an AI company two revenue engines, and card networks treat them differently. A renewal is merchant-initiated: charged on a schedule, against a card the customer authorized once.
A top-up, on the other hand, is customer-initiated: the customer is present, mid-task, and expects it to clear in seconds.
Each needs different transaction data, authentication and retry logic and an AI company needs a payment partner that runs both well, not one built for renewals alone.
One AI customer, seven payment moments
Follow a single AI customer through a year and the difference becomes concrete. Each moment below is a different transaction in the eyes of the card networks, and each fails in its own way when a payment stack treats it as a generic card charge.
The payment layer's job is the right-hand column: sending each charge with the data its card network expects, in every market, so the issuer says yes.
Then there is tax. A merchant of record is a partner that becomes the legal seller of each transaction, taking on tax registration, invoicing and compliance in every market. For a company selling into 55 countries in its first year, it can decide whether a market opens now or next year.
So the question is not only how to charge a subscription but also how to keep every payment flowing in every market you already sell in: renewals that survive expired cards, top-ups that clear in seconds, customers paying in rupees, reais and euros. The companies that answer that best are the ones whose entire business is recurring payments.
Tokens don't pay the bills
Today, tokens are the central currency for companies building with AI. But tokens are what AI companies spend. Their customers pay in dollars, rupees, reais and euros, and every failed payment leaves a token bill paid for nothing. Most AI companies know what a token costs. Few know how many of their payments fail, in which markets, and why. You have optimized what your AI costs to run. Now find out how reliably you get paid for it.
