Stripe currently processes 1.7% of global GDP, and the company was last valued at $159B. In 2025, it processed $1.9 trillion in total payment volume. John Collison is 36 years old, and he founded Stripe with his brother Patrick 16 years ago.

In a rare moment, Cliff from Canva wasn’t the biggest fish in the pond.

Speaking with the Stripe co-founder, Obrecht spoke of Canva as a company re-engineering its business model. During my keynote session on the main stage at Stripe Tour, I took notes for founders not in the room.

For most of Canva's first decade, the cost of serving a free user was effectively zero. That underwrote everything: no watermarks, easy exports, and a free tier that serves millions of people.

"The cost of goods to serve a user has gone from cents per month to now many, many more cents per month," Obrecht said, in a non-specific way. Run a full agentic loop, and the bill compounds: web research, brand lookups, a campaign of designs and images, then intelligence to measure how the content performed once deployed. "All those tokens cost money."

"When we've got sort of 200-plus million free users and say 30-plus million paying users, the paying users need to subsidise the free users, as well as their own generous product."

Canva launched its AI product in April on third-party frontier models. They were "amazing at what they do, but they're also very expensive". The company has since deliberately slowed the rollout while it drove costs down by roughly 90 per cent.

Obrecht said he’s looking to get the “holy trinity” in an AI experience for his customers, made up of "quality, cost and latency", and argues you cannot buy your way to all three. "You can probably get the quality, but the cost and latency you're going to lose." So Canva has in-sourced what it considers core to its customers: image and design generation.

Obrecht was also keen to point out that Canva did not decide to become an AI lab yesterday, but has been making moves in the AI space for over 6 years, initially with the background remover, which led to the acquisition of Austrian AI company Kaleido.

Since then, Canva has also acquired Australian-based AI image-generation startup Leonardo and brought its teams in-house to build out its own models.

On compounding growth. John and Cliff were in agreement that "the majority of that growth comes from three or four key things that work really well for them".

For Canva, it was exposing the template library to SEO, ranking first for template queries, then feeding that search research back into which templates to build. That loop became "essentially another 60 per cent of our growth for a while". Sharing and collaboration came next.

Adding the background remover on Canva was the feature that got John to upgrade to a paid user.

Collison said Stripe found its own version of a growth loop early by building for the rising sharing economy, when it noticed builders cared about "payout logic more than payments". Their customers were building Lyft, Instacart and DoorDash. That observation "very quickly led us to the marketplace functionality", and that business is still growing strong today.

Canva's old model was customer research, a pitch deck with wireframes, a resourcing plan and milestones. "That's kind of the old way. Now it's come to me with the prototype and the smallest team possible to get a first spike out and get customer feedback."

It is faster, requires fewer people and gets closer to the customer sooner. For Canva, it’s now part of the hiring philosophy, where "how do you think about building products" is a filter.

Collison referred to this as the Arctic expedition analogy when they used to get a great galleon loaded with three months of supplies and a library of scientific books, but it was more effective to get two guys and a dog sled.

Obrecht's most concrete suggestion for leaders was to connect AI to as much of your business data as possible.

He’s built a firehose connection into the Canva code repository and expanded it to as many data sources as possible. He then interrogated the business directly using natural language.

He flagged that this has been on his mind for a while. Canva's first attempt at that system, Maestro, failed because the data structures were incompatible: "it was a bit of a mess", and fixing them took real time, but now the AI is so good at interpreting data, it doesn’t really matter.