Canva cut its forecast over AI unit costs

Every AI feature you ship has a per-use cost, and if you don't know what it is, you're running a business you can't price. Canva just published the expensive version of that lesson: in its Q2 CY2026 shareholder update, the company cut 2026 revenue growth guidance from about 30% to about 20% — a one-third haircut — because the AI unit costs behind its newest features outran the money coming in. This is a profitable, $42 billion company with 265 million monthly users. It still got the math wrong.

What actually happened

CEO Melanie Perkins was blunt about the cause. Canva was "relying too heavily on frontier models", several of its own first-party models weren't ready, and — the part that should sting for anyone shipping AI — its "pricing, consumption model and usage controls had not caught up with the outsized demand."

Canva AI 2.0 launched in April 2026 and demand arrived faster than the economics could carry it. Rather than push it to every user, the company held the rollout and spent roughly three months rebuilding the layer underneath. Per B&T, Perkins framed it directly: "Rather than broadly rolling out a product before the underlying economics were ready, we decided to slow the rollout while we rebuilt the architecture, reduced unit costs and strengthened the business model."

The rebuild worked. Canva says the average cost of serving an AI task fell by roughly 90%. Its in-house style-transfer model now runs 23x cheaper than comparable frontier models, image generation 30x cheaper, and video 17x cheaper — built out largely on the back of its 2024 Leonardo.AI acquisition, with 200+ people on AI and 140+ in a dedicated research lab. Q2 revenue still grew 25.2% to US$921.9 million, and the company ended the quarter with US$1.47 billion in cash and nine straight years of profitability. The guidance cut wasn't distress. It was arithmetic.

Why AI unit costs matter for your business

You do not have 140 researchers. That's the point. Canva could afford to eat a quarter of its growth rate, buy an AI lab, and train replacement models. If your AI feature's unit cost quietly exceeds what the customer pays, you get no such runway — you get a support queue and a margin you discover at year end.

Three things are worth copying from this, none of which require a research lab:

Meter before you launch, not after. Cost per task, per customer, per plan tier. If you can't produce that number today for a feature already in production, that's the actual finding.

Frontier models are a starting point, not an architecture. Most production work — classification, extraction, summarization, rewriting — does not need the most expensive model available. Route the cheap work to cheap models and reserve the frontier for what genuinely needs it. Canva's 90% came from doing exactly this at scale.

Usage controls are a product feature. Perkins named them explicitly. Caps, credits, and tier limits aren't hostile to customers; they're what lets you keep offering the feature at all.

The vendor lesson also runs the other way. Canva slowed a rollout to fix its economics — a rational choice that still means a feature you planned around arrives late, gets metered, or moves up a tier. Assume every AI feature in your stack is provisionally priced.

Sources: Startup Daily, B&T.