Canva Halts AI Feature Rollout, Warns Revenue Growth to Slow to 20%; Figma Shares Drop 15%

AI application companies are being tested on whether high inference costs can erode profitability. Design software firm Canva has warned investors that annual revenue growth will slow to 20%, as it proactively halted a planned AI feature rollout that was expected to drive paid subscription growth. Following the announcement, Figma's shares fell about 15% in a single day.

Canva Chief Operating Officer Cliff Obrecht said that before introducing AI, serving a large base of free users was very low-cost, but after deploying AI features, those costs rose sharply, fundamentally changing the unit economics. Reducing AI costs has become more important than ever.

Figma faces a different cost pressure. Chief Financial Officer Praveer Melwani said on an investor call that the company does not currently charge customers for use of its products in the testing phase, so inference costs are borne by the company with no consumption revenue to offset them. This means Figma is using its own funds to subsidize users' use of its AI features while waiting for these tools to move from testing to full commercialization. Figma is also increasing investment in its proprietary AI models, combining internal models with frontier models to support its newly launched AI agents. Training proprietary models requires time and capital investment, and after the guidance release, Figma's shares fell about 15% in a single day.

The cases of the two companies show that product-level appeal has been validated, but the path to converting that into a sustainable profit model is not yet clear. Both companies see proprietary models as the key to breaking through, improving unit economics fundamentally by reducing dependence on costly external models. There is a time gap between the construction cycle of proprietary models and the market's immediate demand for earnings growth. The cost of the AI transition is reflected not only in research and development spending but also embedded in the inference fees behind each user call.