The AI hype cycle just met the income statement. This week, two design-software leaders sent the market a message most AI vendors have been dreading: popularity doesn't pay the inference bill.
Canva warned investors that annual revenue growth would slow to 20% — because it deliberately paused an AI feature rollout that had been driving paid-subscription growth. Demand was too high. Service costs were too high. That combination is the headline. Figma, meanwhile, guided third-quarter revenue growth down from 48% in the June quarter to 36%, admitted its AI tools are still in beta with no clear monetization path, and watched its stock fall roughly 15% in a day.
Two companies, one structural finding: an AI product's adoption curve and its unit economics are now two separate battles — and margin is the one investors are watching.
The Canva paradox: too many users is the problem
Canva's situation is almost ironic. Its COO Cliff Obrecht reportedly told colleagues that before AI, serving millions of free users was "very low" cost. After AI features shipped, "costs climbed dramatically, the unit economics fundamentally changed, and lowering AI cost became far more important than before."
So Canva chose to brake: it slowed its AI rollout rather than buy growth with high-cost compute. That decision directly cut its full-year revenue guidance to 20%. Its bet is in-house models — faster and cheaper than frontier labs for image and video generation — but they missed the rollout window, leaving Canva dependent on pricier external models at the critical moment.
Figma: subsidizing beta features out of pocket
Figma's problem is the flip side. CFO Praveer Melwani put it bluntly: "We don't currently charge customers for features in beta; inference costs are borne by us, with no consumption revenue to offset." Figma is paying out of pocket for every AI invocation while it waits for tools to graduate from beta. That cost structure will press on gross margin — and the market priced it in immediately with a ~15% stock drop.
Figma is also pushing in-house models, mixing them with frontier models to power its new Figma AI agent. But training takes time and money, and public-market patience doesn't.
The framework: every AI call carries a toll
Both cases converge on the same structural point. In the old SaaS model, the marginal cost of serving an extra user was near zero — a gross-margin machine. AI rewrote that equation: now every user interaction carries a real, recurring inference cost. When monetization lags adoption, that toll eats margin. The cost of AI transformation isn't just R&D; it's baked into every single invocation.
This is the first real financial physical exam for the AI application layer. Investors previously priced AI-app vendors on narrative and user growth. Now the unit economics are on the table.
The industry signal
- The AI application layer is moving from growth-at-all-costs to margin discipline — whether vendors like it or not.
- Own-model strategy is now a margin strategy, not a differentiation strategy. Whoever controls inference unit cost controls gross margin.
- The timeline mismatch is the real risk: in-house model build cycles don't match quarterly earnings expectations. Companies will keep eating inference costs until they close that gap.
What to do about it
- If you're building AI features into a product, model inference cost per active user into the unit economics from day one — not as an afterthought.
- If you're buying a stock on AI adoption numbers, ask for the unit economics behind them. Adoption ≠ margin.
- If you're a founder, treat "own your inference stack" as a margin playbook, not a tech flex.
The AI story in software is entering a new chapter: the era of the income statement. Canva and Figma are just the first two companies to get tested.