Software became the best business in the world for one reason above all others.
The second customer cost nothing to serve, and so did the millionth.
That single property is what made freemium possible, what produced gross margins near 90%, and what allowed a generation of companies to hand their product to hundreds of millions of people at no charge in the confident expectation that a small paying minority would cover everything.
Generative AI has broken it, and Canva is the first company large enough and honest enough to say so out loud.
The Australian design company told investors its second-quarter revenue reached $921.9 million, up 25.2%, short of its own guidance and an unusual downgrade for the country's most valuable private technology business. Melanie Perkins, its co-founder, wrote that the average cost of serving an AI task was too high, and that Canva chose to slow its rollout while it rebuilt the architecture and reduced unit costs.
Full-year growth is now expected at around 20%, roughly a third below the ambition set in January.
The detail that matters is not the number but the reason behind it.
Canva has more than 260 million monthly users and most of them pay nothing, which for fifteen years was the company's greatest strategic asset and is now, feature by feature, a bill.
Every prompt run through a large model consumes computing capacity that somebody has to buy, and when the user is free that cost is met from the paying side of the business or from nowhere at all.
Distribution, the thing every software investor was taught to prize above product, has become the liability.
This is not a Canva problem. Figma guided third-quarter growth down to 36% from 48%, conceded that several AI tools sit in testing with no clear commercial path, and watched its shares fall roughly 15% in a day.<cite index="37-1">Its finance chief said plainly that the company does not charge for products in beta and absorbs the inference costs itself.</cite>
Three responses are available and none of them is comfortable.
Meter the usage, as Figma has done with credits, and accept that you have introduced a variable price into a product people bought precisely because it was flat.
Ration the feature, as Canva did, and watch users test whether a competitor will hand them the same thing free.
Or build your own models, which is the route Canva says it is taking, and in doing so march down the technology stack at the exact moment OpenAI and Anthropic are marching up it.
That last option deserves more scepticism than it gets, because it commits a design software company to a capital and research programme it did not choose and cannot easily abandon.
There is a broader reckoning implied here for anyone who has looked at the revenue running into the model providers and concluded the application layer must be thriving too.
A meaningful share of that revenue is being paid by companies serving it to users at a loss, funded by venture capital, by profitable legacy products, or by both.
Canva is thirteen years old, profitable, and has one of the strongest consumer brands in software.
If the arithmetic is flinching there, it is not working anywhere in the cohort of AI-native startups whose entire promise rests on giving the product away first.
The industry has spent three years arguing about the cost of training models.
The cost of serving them is what will decide who survives.
