Canva's $10 Billion Markdown: Blackbird and Airtree Cut Valuation as AI Costs Bite

Canva, the Australian design-software giant long treated as the exception to every startup rule — profitable, fast-growing, beloved — has just received the most sobering markdown of the AI boom. On August 13, Blackbird Ventures and Airtree, two of its longest-standing backers, wiped US$7.1 billion (A$10 billion) from its valuation, marking the company down to US$34.9 billion from the $42 billion struck at an employee share sale in August 2025, following a fresh independent valuation. Canva’s own internal valuation has reportedly fallen to $31 billion, down from $38.9 billion.

The proximate cause is a confession few companies of this scale have made publicly: AI features turned out to be dramatically more expensive to deliver than anyone modeled.

The numbers behind the markdown

The trouble surfaced in early August, when The Information first reported that Canva had cut its 2026 revenue growth forecast from 30% to 20% — a ten-point haircut on a company doing roughly $4 billion in ARR. Second-quarter revenue came in at $921.9 million, up 25.2% year-on-year — a strong number in isolation, but a miss against the company’s own guidance, with some AI features running as much as six months behind schedule.

In her Q2 letter to shareholders, co-founder and CEO Melanie Perkins was unusually direct about what went wrong. The company, she wrote, had been “relying too heavily on frontier models” — routing AI traffic through expensive third-party models (the API bills of OpenAI and peers) because Canva’s own first-party models weren’t ready, while pricing and usage controls hadn’t caught up with demand. Usage of Canva’s AI suite grew faster than management planned, and each generation, edit, and magic-wand click burned frontier-model tokens at a cost the subscription pricing never contemplated.

The context makes this sting. This is not a venture-funded startup burning cash on inference — it’s a mature, profitable platform valued at $42 billion with more than $1 billion in cash, more than 260 million monthly users, and 31 million+ paying subscribers, admitting it got its AI cost model wrong at scale. Fortune’s assessment was blunt: the AI that was supposed to power Canva’s next act “also exposed software’s” underlying economics.

The fix: rebuild, route, and reprice

The interesting part of the story is the response. Rather than shipping AI features at unit economics it couldn’t sustain, Canva chose to slow the rollout and rebuild. Over the past three months, the company re-architected its AI infrastructure and moved more inference onto its own models, and now claims a 90% reduction in the cost of serving an AI task. If that holds as usage scales, it’s a structural change in unit economics rather than a one-quarter accounting trick.

The new architecture is a textbook exercise in cost-tiering: free-tier users get routed to Canva’s own cheaper models, while frontier-model access is reserved for paid tiers where the unit economics can support it. And for the heaviest users, Canva introduced AI Pass at US$100 per user per month — an add-on stacked on top of Pro or Business that isolates heavy AI consumption into its own price point, insulating the core subscription from inference costs.

Missing from the briefing, as analysts have noted: AI Pass adoption figures, gross margin on AI usage, and how much of the guidance miss traces to the $100 tier landing softly versus features simply arriving late. Canva is private, so these numbers arrive via shareholder letters as reported by the financial press, not public filings — which makes the internal discipline more notable, not less. Nobody forced this disclosure; the market wasn’t punishing the stock because there is no stock.

The ChatGPT double-edge

The competitive backdrop adds a layer the markdown numbers don’t capture. Canva’s relationship with ChatGPT is genuinely double-edged. The company reported more than 26 million conversations with its ChatGPT app by October 2025 and claims to be a top-10 referred domain from ChatGPT — real distribution, at real scale. But that same channel lets users generate a design through ChatGPT without ever opening Canva’s app, and OpenAI’s own creative tools keep improving. The AI that inflates Canva’s costs is sold by the same company whose chatbot is becoming an alternative front door to design work.

This is the squeeze every application-layer SaaS company now lives in: frontier-model providers set the cost floor, agentic assistants threaten to become the interface, and the application vendor must somehow fund the AI bill while defending the customer relationship.

A warning shot for the whole SaaS sector

The SaaS CFO community has already adopted Canva as the cautionary tale of 2026. The lesson drawn most often: the difference between companies managing AI costs and companies being managed by them is whether AI unit economics are tracked as a distinct line — cost per user, per task, per outcome — or buried inside COGS until the number gets big enough to move guidance. Canva, the analysis goes, got caught in the AI learning curve; Wix engineered its way out (its Base44 AI builder went from near-zero gross margin to about 60% in half a year by building its own model, Base 1) precisely because it tracked the numbers early.

For Canva’s IPO ambitions — Blackbird told LPs late last year the company was “ready” for a 2026 listing — the markdown is a complication, not a cancellation. A 17% valuation cut from sympathetic insiders, on the strength of an honest cost reset and a claimed 90% unit-cost improvement, is survivable. But it reframes the listing narrative: Canva will go public not as the effortless growth story of the last decade, but as a case study in whether application-layer software can own its AI cost curve.

That question — can anyone below the model layer make the AI math work? — is now the sector’s central accounting problem. Canva just gave investors the first large-scale datapoint, and the market’s first answer was minus ten billion dollars.