Canva Cut Its Own Forecast to Fix AI Costs

The Forecast Canva Chose to Cut

When Canva reported its second-quarter 2026 results, chief executive Melanie Perkins opened with the number analysts were least expecting: full-year growth guidance, cut from 30 percent to 20 percent, a full third off the company's own prior target. She did not present it as a miss. She called it "a deliberate decision to get the economics right" on the cost of running Canva's AI features - the kind of language a chief executive uses when a number moved because leadership chose to move it, not because the market forced it.

The quarter itself was not weak by ordinary standards. Revenue reached 921.9 million US dollars (1.32 billion Australian dollars), up 25.2 percent year on year, a growth rate most software companies would report with pride. It simply fell short of the pace Canva's own roadmap had assumed, and some of its newer AI features shipped as much as six months later than planned - a delay that, in hindsight, traces to the same underlying problem as the forecast cut.

Nothing about this reads like a company in trouble. It reads like a company that priced its own AI ambition against a number that turned out not to hold, noticed early enough to matter, and went looking for why.

Why the Economics Stopped Adding Up

The diagnosis: Perkins named three specific causes, and each one is worth reading as a checklist rather than an excuse. The average cost of serving a single AI task inside Canva's products had climbed too high to support the pricing built around it. The company had been, in her words, relying too heavily on frontier models - the large third-party systems from providers such as OpenAI - for work that did not need frontier-grade capability to do it. And Canva's own pricing, consumption model and usage controls had not caught up with how much demand actually showed up once the features were live at scale, rather than in a pilot.

The stakes: each of those failures looks small in isolation and compounds badly together. A feature costed at pilot volume looks affordable. The same feature, multiplied by millions of users who now expect it bundled into a subscription rather than metered per use, turns into a cost line that grows faster than the revenue line sitting next to it. Canva builds products for design teams, marketers and small businesses across Europe who feel this same mechanism from the buyer side every time a vendor's AI add-on price moves without warning; this is the supplier side of exactly that mechanism, run by a company large enough to publish the numbers.

What makes the diagnosis unusually useful is that it is not a story about AI being expensive in the abstract. It is a story about a specific mismatch: paying a frontier provider's list price for tasks that a smaller, purpose-built model could handle at a fraction of the cost, at a volume where that fraction compounds into real money.

Ninety Percent Cheaper, Built In-House

The fix: Canva spent roughly three months rebuilding its AI stack around models it owns, built substantially on the technology behind Leonardo.AI, the Sydney-based generative-AI startup Canva acquired in 2024. The logic is narrow and specific: keep frontier models for the tasks that genuinely need frontier capability, and move the high-volume, repeatable inference - the bulk of what a design tool actually asks an AI model to do - onto infrastructure Canva controls and prices itself.

The results are the part worth sitting with. Canva reports a roughly 90 percent cut in per-task AI serving cost overall. Broken down by model, the company's in-house image model came out about 30 times cheaper than comparable frontier models, its video model about 17 times cheaper, and a style-transfer model about 23 times cheaper. Those are not marginal efficiency gains from better prompting or caching. They are the difference between a business model that works at the volume Canva actually serves and one that does not.

None of this reads as a company under financial pressure. Canva remains profitable for a ninth consecutive year and holds 1.47 billion US dollars in cash. The correction was a choice made from a position of strength, not a rescue made from a position of weakness - which is exactly why it is worth studying rather than filing away as a distress story.

The Build-Versus-Buy Trap Any Roadmap Can Fall Into

The pattern: Canva's reversal is a live, numbers-backed case of a trap that catches far smaller companies with far less room to absorb it. A product roadmap gets scaled against a third-party frontier model's list price before anyone has modelled the unit economics at real usage volume. The pilot pencils out. The rollout does not, because consumption outpaces what the vendor's own pricing and usage controls were built to absorb - and by the time that becomes visible, the roadmap, and the customer expectations built on it, are already committed.

What to do differently: treat frontier-model API cost as a variable that can move faster than your own product pricing can react to, not as a fixed input you price around once and forget. Before shipping an AI feature at scale, model the cost per task at something like ten times your expected volume, not at pilot volume, because pilot volume is precisely the number that hides this problem until it is expensive to fix. Canva's specific answer - bringing the highest-volume, most commoditised inference in-house and reserving frontier APIs for the genuinely hard tasks - is a concrete pattern any product team with the resources to build or buy a smaller model can copy; the modelling discipline behind it is available to every team regardless of size.

The lesson does not require a company to build its own AI models to be useful. It requires treating vendor pricing as a moving part of the business model, tested against the volume the product will actually see, before that volume arrives and the bill does the deciding for you.

Read next: Your AI Vendor May Get a Government Kill Switch  |  Rust Draws the Line AI Cannot Cross