Token economics & why the payments companies want to be your AI bill
Over the last month, three software companies told investors that AI was eating their margins. Then Stripe and Ramp both bought into the layer that fixes it. Here’s an explainer on how it’s all connected, with models you can play with.
Rethinking software’s oldest assumption
For over twenty years SaaS has operated under the model that serving one more customer cost almost nothing. Inference changes that.
Write the code once, and the ten-thousandth user is nearly free. That single fact is why software companies are worth what they’re worth and why VCs are in business.
However, a funny thing’s happened. Over the last few years we’ve seen everyone run to take advantage of the value AI-enabled features can bring to their products applying the same software thinking to this.
What’s becoming apparent though is that it’s looking like AI breaks this. Every time someone clicks a button that runs an AI feature, the company pays for it with a small amount of computing power bought from a model provider in real time. The industry calls this inference. Think of it as the electricity bill for thinking.
Individually those charges are tiny, fractions of a cent. But they scale with usage and in early August, three companies got the bill and it wasn’t what they were hoping to see.
Three companies, one problem. Figma is public, so its numbers are audited: revenue grew 46% while gross margin fell from 91.5% to 79.6% in four quarters. Canva is private and cut its growth forecast from 30% to 20%. Wix went the other way, taking an AI product from roughly no margin to about 60% in six months. Canva and Wix figures are self-reported.
A word on gross margin, since everything below revolves on this. It’s what is left of each dollar of revenue after paying the direct cost of serving that customer. A 90% gross margin means ninety cents of every dollar is available for salaries, marketing, R&D and profit. Software has historically lived in the high eighties or low nineties. Supermarkets live around 25%. It is the number that most explains why investors value software the way they do.
Figma’s is the clearest case because it is a public company and has to show its working. Revenue grew 46% year on year. Cost of revenue grew roughly 250%. On the earnings call, CFO Praveer Melwani attributed the compression to broader and deeper adoption of AI features, with users reaching for more capable, meaning more expensive, models.
Canva’s version was a little more dramatic. The company cut its 2026 growth forecast by a third. The reason was not weak demand but more the opposite. Users of Canva’s new AI tools made roughly three times as many designs as the company had planned for. Canva got punished for building something people liked.
Now here’s the kicker. Nothing went wrong at any of these companies. The product worked better than intended but that’s what caused the problem.
The success penalty
Engagement used to be pure upside. AI features have their downside.
In traditional software, a customer who uses the product every day costs no more than one who logs in monthly, but is far likelier to renew. The chart below draws out what changes when engagement carries a marginal cost.
Move the sliders and watch what happens. The horizontal axis is how much a typical user leans on the AI features. The vertical axis is gross margin. Every step to the right is a step down.
How to read it. The line starts at whatever you set “gross margin before AI” to and it defaults to 91.5%, Figma’s disclosed margin before AI adoption scaled. The horizontal dotted line marks 79.6%, where Figma actually landed four quarters later. Both endpoints come from SEC filings. Everything between them is this model, not disclosure.
One slider behaves differently from the rest and is an illustration of the problem at hand. Tasks per user doesn’t change the shape of the lines it just walks you along them. It is the only input here that is not a property of your business. It is a property of how much your customers enjoy the feature. You do not control it, and the direction hurts.
Everything else bends the curve. More users or higher adoption makes the fall steeper. A more expensive model makes it steeper. Growing revenue makes it flatter, because the same dollar cost is a smaller slice of a bigger pie. That last one is real defence, and it’s the thing Canva didn’t get. Usage tripled but revenue didn’t.
A note on this model
It is a visual tool, not a valuation. The two endpoints on the chart are audited public figures and the curve between them is arithmetic built on stated assumptions. It is meant to show the shape of the problem, how quickly margin falls, and what actually slows the fall, not to predict any particular company’s numbers.
What routing is, in plain terms
The majority of AI assisted tasks don’t need an expensive model. Routing is the plumbing that helps you decide.
Every company facing this problem is currently reaching for a fix with a previously unglamorous name: routing.
To sort a support ticket into a category, pull a date out of an invoice, or summarise a paragraph, a cheap, small model handles these taks perfectly well. Only a minority of requests genuinely need the frontier of the newest, largest, priciest model available.
A router is a piece of plumbing that sits between an application and all the model providers. Developers connect to it once. Behind it, each individual request gets sent to the cheapest model that still clears a quality bar the company sets.
Two comparisons usually land. The first is payment processing itself – Stripe already decides, invisibly and per transaction, which network and method to send a card payment through to get the best result at the lowest cost. Same shape, different commodity. The second is electricity: you do not pick a power station every time you boil a kettle. Something upstream decides, based on price and load.
There is a catch, and it is the whole ballgame. Routing only works if you can tell whether the cheap answer was good enough. Measuring this reliably is a genuinely unsolved problem. Part of the reason is that you can’t rely on users to determine this for you as most companies have chosen to do. Two things sit in the way. Firstly, we all an inflated sense of the importance of our tasks. If I’m speaking to a founder for an hour, I’ll tell myself I most certainly need the top models to summarize the call. This is even more so relevant to the Canva case that if I’m appropriating budget from hiring a designer to create something for me, I need to squeeze the best of the tool as I can. And secondly, where the real issue is, if I as a user don’t see the bill, why do I care?
This is where the real difficulty and probably the real defensibility and cost protection sits.
What routing actually buys
Routing flattens the curve. It does not repeal it.
The two sliders below control the same model as before. The first sets what share of requests go to an expensive frontier model. The second sets how much cheaper the alternative is – either a smaller off-the-shelf model or one the company trained itself.
Two versions of the same business. The top bar sends every request to an expensive model. The bottom bar uses whatever mix you have set. The vertical line marks one dollar of revenue — when the inference band pushes past it, costs exceed revenue.
The headline claims in the market sit inside this picture. Ramp says its customers cut inference spend 40% on average. Canva says it took roughly 90% off its cost per task, by combining routing with models it trained itself – reporting an in-house video model 17 times cheaper than the frontier equivalent, and an image model 30 times cheaper. Both figures come from the companies themselves; nobody independent has checked either.
Dial the sliders to those levels and something shows up. A 40% cut in inference spend is real money, but it is usually worth fewer margin points than it sounds – because inference is only one part of the cost base. Routing buys time.
That is also the honest reading of the chart in section II. Turning routing on does not lift the line so much as flatten it. The penalty for success gets gentler, which is what lets a company grow into it rather than get caught by it.
So the payments companies bought the meter
Within two days, Stripe and Ramp both took a position in the layer that meters the tokens.
On August 19th, Stripe announced it had agreed to acquire OpenRouter, the largest independent routing platform with a single connection point to over 400 models from more than 80 providers. Reported prices differ by outlet: Bloomberg said more than $7bn, the New York Times about $7.5bn, Axios more than $8bn. At the time of writing, neither company confirmed a figure. OpenRouter had raised at a $1.3bn valuation in May, three months earlier.
The next day, Ramp launched Router.com, a routing infrastructure it had built for itself three years ago and now offers to everyone, free through the end of 2026.
Try this. Set spend to about $2.5bn and the fee to 5.5%, and you land near OpenRouter’s reported annualised revenue and a multiple around 50 times. Then drag the fee toward zero. That is Ramp, giving routing away free through 2026.
So why payments companies, of all people? Well, taking a guess it seems to come down to four reasons that compound.
They could see it coming before anyone else. Ramp watches AI charges land on corporate cards and invoices. Its own index puts business AI spend at 20.7 times its June 2025 level. On the other hand Stripe processes payments for a large share of the AI industry and was already OpenRouter’s own payment processor. It watched that revenue curve in its own billing data before it bid.
Taking a small fee on flowing volume is the business model. A gateway skims a percentage of the compute passing through it. Structurally that’s a payments income statement pointed at a new kind of volume.
Choosing a model is becoming a finance decision. When inference was 2% of costs it was an engineering choice. When it moves a public company’s gross margin by twelve points, finance will poke their head in to take a look. Plus the finance team already uses these vendors.
They did’nt have to build an audience. Neither company needed to win developers from scratch. They already own the person newly being asked to approve the bill.
The money in this cycle probably won’t go to whoever builds the best model but to whoever meters the tokens. This is a position that gets more valuable as models become interchangeable, not less.
Figma’s CFO described this playbook of routing by task complexity, staying model-agnostic, and training first-party models on an earnings call in May. Stripe bought the routing layer in August. The buyers were not spotting something nobody had noticed. They were buying into a need their customers had already announced in public.
We could be wrong about this
With everything going on today, we know from experience that sometimes our predictions dont work out. Routing may turn out to be a feature rather than a category.
We would rather put the counter-argument in our own words than wait for someone else to. There is a real chance these deals will look expensive in hindsight if the following continue to be true.
- 01
Everyone already ships one
Amazon, Google and Microsoft all offer multi-model access on their clouds. There is a popular free open-source option. Cloudflare and Vercel ship gateways too. It is not obvious this layer stays independent long enough to be worth billions.
- 02
The fee is already falling
Ramp is giving Router away free through 2026 and has not said what it will cost in 2027. Free-at-launch is often what a layer looks like just before it stops being able to charge for itself. Drag the fee slider above to zero to see what that does.
- 03
Convergence kills the value of choosing
Routing is worth something because models differ sharply on price, speed and quality. If the leading models converge, the value of picking between them drops toward nothing.
- 04
Not everyone will route their data
A request has to reach the gateway to be routed. Router’s stated default keeps inputs, outputs and tool calls for a year unless you opt out. Plenty of companies simply will not send sensitive material through a third party.
There is a fifth wrinkle in our thinking here. Routing assumes cheap models stay cheap. In early August, DeepSeek who have been notorious for pricing aggressively and arguably started the price war for tokens warned of a significant price increase and floated higher rates during peak business hours, apparently because demand had outrun the computing power available. Prices per unit have fallen hard for two years, and there’s reason to believe that tokens are as expensive as they’ll ever be. But prices could go up.
What we’re watching in and out of our portfolio
Three questions have moved to the top of our list for anything AI-native.
What does it cost you to serve one AI task? Not blended cost of revenue but the specific figure. If a founder can’t answer this, they are exposing the same blind spot Canva just disclosed publicly.
What is the gross margin on the AI feature itself? A healthy blended number can hide a feature that loses money on every use, but as seen above that gap widens as the feature succeeds.
What share of your requests go to an expensive model, and what is the plan to move it? The answer tells you how much headroom is left before the cost problem becomes a pricing problem.
On where we are looking for companies, well the routing layer itself now has Stripe in it, which is a hard fight to pick. The adjacent problems look more interesting. Deciding whether a cheap answer was good enough is unsolved, and difficult. “Taste is the moat” as has become popular parlance in tech circles. But how do you measure it? So is showing a finance team where token spend actually goes. It seems logical that plenty of companies will want to follow Wix and Canva in training smaller in-house models without building a research team to do it.
That all said, the broader point we’ve been trying to make is simple. For years, the cost of serving a software customer was close enough to zero that nobody modelled it. A lot of business models were built on top of it and that assumption is now wrong.
Sources
Where figures are self-reported by a company rather than audited or independently verified, we have said so. Deal values for the Stripe transaction differ materially between outlets; neither party confirmed a price.
Figma
- Figma Announces First Quarter 2026 Financial Results – Q1 revenue of $333.4m, up 46% year over year.
- Q1 2026 earnings call prepared remarks (PDF) – CFO commentary on gross margin and AI adoption.
- Figma Q1 2026 press release as filed
- Earnings call coverage – routing, model-agnostic architecture and first-party models named as margin levers.
- GAAP gross margin of 79.6%, down from 91.5% a year earlier
Canva and Wix
- Canva cuts 2026 growth guidance from 30% to 20% – discussion of the AI cost driver. Canva is private; figures are company-reported.
- Canva revenue, users and AI usage estimates – 265m monthly active users; ~$4bn ARR at end-2025.
Stripe and OpenRouter
- Stripe agrees to acquire OpenRouter – the official announcement. No price disclosed.
- Deal finalised at more than $7 billion
- New York Times reported figure of about $7.5 billion
- More than $8 billion in cash and stock
- OpenRouter’s $113m Series B at a $1.3bn valuation, May 2026
- Deal coverage and founder commentary
Ramp and Router.com
- Ramp launches Router.com – 40% average customer saving and the 20.7x AI Index figure, both Ramp-reported.
- Product site – pricing, coverage and free-through-2026 terms.
- Analysis noting neither the 30% nor 40% figure has been independently audited
- Limits of the offer – US only, no 2027 pricing announced
Written by the team at White Star Capital. Nothing here is investment advice. The interactive models are illustrative and built from public inputs; only their labelled endpoints are drawn from company disclosure.
