Canva cut 2026 growth guidance from 30% to 20% because it was "relying too heavily on frontier models"

Canva reported Q2 2026 revenue of $921.9M, up 25.2%. Then it cut full-year growth guidance from 30% to 20%. CEO Melanie Perkins attributed the cut to routing too much AI traffic through expensive third-party frontier models while Canva's own models weren't ready. The expected H2 2026 listing now appears to slip into 2027. Fortune

Growth rate. Not margin. That distinction is the whole story.

Everyone accepted a while ago that AI features compress gross margin. That's a known tax, priced in, discussed on every earnings call since 2024. What Canva disclosed is different: inference cost got large enough to change how fast the company grows and to delay a public offering. When your cost of goods sold sets your growth ceiling, it stopped being a finance-department problem and became a product-strategy problem.

Canva says it has since cut AI servicing cost per task by about 90%, by building proprietary models and buying AI startups. Note the unit they chose. Cost per task. Not cost per user, not cost per month.

Wix ran the same play and got there first. It guided Base44 non-GAAP gross margin to about 60% for H2 2026, up from essentially zero entering the year, crediting Base 1, its own LLM. Wix says Base 1 produces better results for Base44's specific workflows while materially lowering inference cost versus leading external providers, and it plans to reinvest the entire gross-margin benefit into Base44 sales and marketing. Wix

That's the playbook, stated plainly: a narrow in-house model tuned for one workflow beats a general frontier model on unit economics, and the recovered margin becomes distribution budget. Not "train your own foundation model." One workflow, one model, one cost line you control.

The pattern showed up in four unrelated categories within ten days. Canva in design. Wix in website building. Atlassian holding about 89% non-GAAP gross margin while carrying 5M monthly Rovo users at 20%+ month-over-month credit growth. And Databricks, where Ali Ghodsi named the "margin bill for agents" outright. The SaaS CFO

Ghodsi's version has a twist worth stealing. Databricks crossed $7B ARR growing 80%, a 30-point acceleration, at a $190B valuation. His observation is that agents generate far more queries than humans do, which compresses gross margin. But Databricks is consumption-priced, so it gets paid for the extra volume. A seat-priced vendor eats it. SaaStr

If you sell software with an AI feature and you price per seat, run that math this week. And if you can't say what one task costs you, you have the blind spot Canva just disclosed at public-market scale.