Canva builds its own AI model after a rented one broke its margin
Research headcount tripled in a year
What happened: Canva launched Canva Craft, an in-house AI image model trained on its own templates, on 23 October. Per AFR, the company says it is significantly cheaper to run than the third-party models it replaces, though it gave no cost figures. The same reporting says frontier-model token costs forced Canva to delay its AI 2.0 launch in August, issue a rare revenue downgrade, and take a $10bn cut to its valuation, to $50bn. Canva's research headcount has tripled over the past year.
Why it matters: A company large enough to feel a token-cost problem in its valuation has concluded that building beats renting. Any business whose product margin depends on a third-party model's price list now has a live example of what happens when that price moves, and of the alternative.
Signal:Model choice is a gross-margin decision before it's an architecture one. Canva's own numbers make that explicit: a vendor's pricing moved, the revenue line and the valuation followed, and the response was to build the capability in-house rather than keep paying the toll. Boards whose products are wrapped around someone else's model are carrying that same exposure, usually without a number attached to it yet.