Canva Cut Its Growth Forecast by a Third to Pay Its AI Bill

Canva told shareholders in early August that it now expects to grow about 20% this year rather than the 30% it planned in January. The thing that took the other third was its own AI features.

Revenue for the June quarter was US$921.9 million, up 25.2% on the year before. CEO Melanie Perkins told shareholders that the company had shipped AI faster than its own economics could carry it:

“Several of our first-party models were not yet ready for release, and our pricing, consumption model and usage controls had not caught up with the outsized demand we were seeing.”

So Canva slowed down on purpose.

“We decided to slow the rollout while we rebuilt the architecture, reduced unit costs and strengthened the business model.”

The fix wasn’t a price rise. Canva changed the engine.

Until this year the company was renting frontier models, the large general-purpose systems that OpenAI and Anthropic sell by the unit of work. Every time somebody clicked an AI button inside Canva, Canva paid someone else. It has since rebuilt around models it owns, and says serving one AI task now costs nearly 90% less than it did in April. Its video model runs about 17 times cheaper than a comparable frontier model. Its image model, about 30 times cheaper.

Nobody sends the customer a changelog when that swap happens.

What a 30x cheaper model means on your side of the screen

You won’t see a line item. You’ll see outputs that feel a little different, and an invoice that doesn’t move.

That is the shape this cost problem takes for the buyer. The vendor absorbs the bill for a while, then re-engineers underneath the same button, and the quality question lands on the marketing team that has already built a workflow on top of it. If your social templates, your ad variants, or your first-draft imagery come out of a tool with AI baked into the seat price, the model behind that seat is now a moving part the vendor is actively motivated to make cheaper.

Derek Hernandez, senior research analyst at PitchBook covering where SaaS and AI meet, told Fortune what breaks here:

“AI is making SaaS no longer a zero marginal cost solution, which has really been what I would call a lot of software’s secret sauce up until now.”

That sentence is the whole story. Traditional software cost roughly the same to serve the ten-thousandth customer as the tenth, which is why seat pricing worked and why vendors could keep adding features without repricing. An AI feature bills every single time it runs. Heavy users now cost real money, and a flat seat price stops covering them.

Canva isn’t an outlier, it’s just the one that published a number. Figma’s free cash flow margin fell to 14% in the second quarter from 27% in the first, and it guided to 36% revenue growth for the third quarter against 48% in the second. Two of the tools sitting on most marketing teams’ desks reported the same squeeze in the same month. We covered the design side of that stack when Figma’s own research showed product teams dropping the handoff.

The bill was always going to arrive somewhere

Marketing teams have spent two years watching their own AI spend get harder to explain, and the conversation has mostly been about the tools bought deliberately: the seats, the credits, the pilot that never got an owner. We’ve written about what happens when the AI budget starts getting rationed and about compute costs pushing into marketing budgets.

This is the other direction. The AI you didn’t buy, inside the tool you did, priced into a seat that was quoted before anyone knew what inference cost.

Vendors have three levers and only three. Raise the price, which is visible and gets renegotiated. Meter the usage, which is what credits and AI allowances are, and which shifts the risk to you. Or make the model cheaper, which is invisible and is what Canva chose. Perkins framed the whole exercise as a deliberate decision to get the economics right, and from a shareholder’s seat it plainly was. From a user’s seat it means the thing you tested in March isn’t necessarily the thing running in September.

None of this is a reason to rip anything out. Canva still grew 25% in the quarter and is sitting on US$1.47 billion in cash. The company is fine. The pricing model everybody quoted you on is what’s under strain.

What follows is a habit rather than a project. When a vendor tells you it has made its AI dramatically cheaper to run, treat that as a product change and not a press release, because it’s one. Re-run the handful of prompts your team actually depends on, keep the outputs, and compare them next quarter. If the tool got worse, you’ll have evidence instead of a feeling. If it didn’t, you’ve lost 20 minutes.

Write down the 30x rather than the growth cut. A vendor that can make its image model 30 times cheaper to run had 30 times the headroom in there all along, and so does everyone still renting the frontier models. The repricing of your stack has started. The first round of it is happening behind buttons nobody is going to announce.