Key Takeaways

  • SaaS products historically scaled on near-zero marginal costs, but AI generation turns zero-dollar actions into direct expenses measured in cents per prompt.
  • Canva balances over 200 million free users against 30 million paid subscribers, meaning paid revenue must subsidize free usage without destroying gross margins.
  • Canva slashed operational inference costs by 90% by acquiring Leonardo and training proprietary image generation models in-house.
  • Text LLMs are fungible commodities you can rent, but core domain models (like visual design for Canva) must be owned to survive.

The Death of Zero Marginal Cost

For two decades, software economics ran on a simple rule: serving user number 10,000,001 cost basically the same as serving user number 10,000,000. That assumption built the modern freemium playbook. You could give away generous free tiers because database writes and static asset hosting cost fractions of a microcent.

Generative AI erased that safety net overnight.

As Canva COO Cliff Obrecht explained to Stripe president John Collison on Cheeky Pint, adding AI to visual creation altered their underlying math. “The cost to create a design before, which was essentially nothing, has now turned into cents,” Obrecht noted. “And that at hundreds of millions of users scale really has an effect on the business that we've had to adapt to.”

Canva supports more than 200 million free users alongside 30 million paying subscribers. When every free user prompt rings up a meter on a third-party server, free tiers stop being top-of-funnel marketing. They become an existential burn rate. Paid users cannot generate enough margin to subsidize a runaway API bill from OpenAI or Anthropic.

The Holy Trinity of In-House Models

To survive at scale, Canva stopped treating AI as an external API dependency. Instead, they built an internal research team and acquired the Australian generative AI startup Leonardo to train proprietary image and design models.

The result was a 90% reduction in inference expenses. Moving workloads to custom internal models allowed Canva to tune performance specifically for their canvas interface rather than paying for general-purpose compute.

Obrecht points to three constraints that dictate their technical roadmap: “It's the combination of quality, cost, and latency that is the kind of holy trinity.” If a model takes ten seconds to generate an asset, users leave. If it costs four cents per run, the business dies. If it looks terrible, nobody cares how fast or cheap it was.

Canva also makes a sharp distinction between text and visuals. Large language models for copy editing or text generation remain fungible commodities. Canva swaps between them freely based on provider pricing. But visual design represents Canva's primary moat.

“And only by building your own research team and training your own models, particularly for image and design,” Obrecht explained, “we really needed to be in control of our own destiny there.”

What to Do With This

Audit your product's AI features by calculating the gross margin per user action. Group your prompts into commoditized tasks (like text summarization) versus core domain workflows (the single output your product is bought for). If your core workflow relies on a frontier model API, start benchmarking smaller open-source models (like Llama or Mistral variants) on fine-tuned domain data this week. Set a target to run your high-frequency workflows on self-hosted infrastructure before your user base scales ten-fold.