Canva's Truth Tax

Last week, Canva deleted nearly eight billion dollars from its own valuation.

Nobody made it happen. No regulator demanded it, no short-seller forced it, no market printed it. Canva is private, profitable and under no obligation to tell anyone anything. It commissioned an independent assessment of its own worth and cut the number employees can sell shares at from $38.9 billion to $31 billion. Its two longest-standing investors, Blackbird and AirTree, cut their own mark from $42 billion to $34.9 billion.

The company marked itself down further than its investors did. That action is as unique as this Aussie unicorn is.

I've been writing about this company for a while. In January last year, when Canva was valued at $26 billion, I argued that its true innovation was like the winged keel of Australia II: hidden beneath the waterline, invisible to competitors until the crucial moment. In April this year, in The Human Canvas, I argued that Canva had built the most behaviour-aligned creative platform in the market and that its biggest risk was its own marketing: fix the brand before the market catches up with what you've built.

The market has now caught up with something. And what Canva did when it arrived is one of the most instructive leadership stories of the year.

Before I make the case for what Canva got right, let me make the case against, because it's real and a sharp reader will already be assembling it.

A growth forecast cut by a third is a miss, not a virtue. Canva told investors to expect 30% revenue growth this year; it now expects 20%. Second-quarter revenue came in at $921.9 million, up 25.2%, a number most software companies would frame and hang on the wall, but one that missed Canva's own internal guidance. "A deliberate decision to get the economics right" is the framing every management team reaches for after the numbers have already moved. The Australian Financial Review, which broke the story, chose a blunter headline register: AI bill shock.

Blackbird and AirTree didn't mark the company down out of admiration for its character. Their external valuers did it because the discounted cash flows demanded it, and because both firms answer to their own investors, who will one day want to see this position sold at a defensible price. Fiduciary duty produced that 17%, not sentiment.

A cynic reads the whole episode as a company discovering, expensively, that the AI features it shipped cost more than the revenue they bring in, and dressing the discovery in the language of intentionality.

Most of the great corporate AI adoption story of the past three years is a text box.

Chat interfaces. Copilots docked in the corner of software that works exactly as it did before. Demo-day videos of three-second cats riding skateboards. Visible AI (the kind you can screenshot for the board pack) is everywhere, because visible AI is cheap. A chatbot answering occasional questions costs pennies per user. It photographs well, it fills a slide, and it never troubles the P&L.

Canva did the harder thing. It dissolved AI into the product so completely that the user never meets a prompt. Canva AI 2.0, launched in April, takes a plain idea typed by someone with no design training and returns a fully layered, editable, on-brand campaign with the model selection, the prompt engineering and the orchestration all happening below the waterline, exactly where the winged keel lives.

This is the democratisation thesis extended one full generation. Fourteen years ago, Canva removed the requirement to be a designer. Now it is removing the requirement to be a prompt engineer, the new gatekeeping skill AI quietly installed just as it demolished the old ones.

And here is the tell that separates real adoption from theatre: usage tripled. Demand for the new AI features, in CEO Melanie Perkins' words, "significantly exceeded" the company's expectations. Invisible AI gets used. Visible AI gets demoed.

Which brings us to the bill, because the adoption story and the honesty story are the same discipline, applied twice: treating AI as infrastructure rather than theatre. Companies that adopt theatrically never get a compute bill big enough to require honesty. Canva's markdown is, perversely, a certificate of authenticity. It hit the cost wall first because it adopted first, deepest, and at a scale of one billion designs a month.

The mechanism of the wall is worth one plain paragraph, because it rewrites the economics of every piece of software you're paying for. Traditional SaaS had a secret sauce: serving one more user cost approximately nothing. Build once, sell infinitely, watch the margin compound. AI inference breaks that. Every AI-assisted design, every generated video, every agentic task arrives with a compute bill attached. The more successful your AI features, the bigger your bill. Success itself became a cost centre.

Canva's response is the technical spine of this story, and it deserves more attention than the markdown got.

Perkins' team spent three months rebuilding the architecture. Not pausing AI but re-plumbing it. Canva now builds its own specialist models and routes work between them and the frontier labs depending on the task. The numbers she has put on the record are remarkable: Canva's style-transfer model now runs 23 times cheaper than frontier alternatives, its image model 30 times cheaper, its video model 17 times cheaper. Add usage-based pricing catching up with consumption, and the result Perkins reports is a near-90% reduction in the cost of serving a single AI task since April, even as usage tripled.

Notice what that architecture actually is. It is not frontier-versus-open-source, the debate the industry keeps having. It is a third position: frontier intelligence as a premium ingredient, not a default. A router that sends a billion monthly tasks to the cheapest model that clears the quality bar, and reserves the expensive calls for the moments that genuinely need them. Orchestration, the thing I wrote a book about, applied to the model layer itself.

If you are a CMO signing an AI-embedded software renewal this quarter, this hands you a new question for the vendor, and it is not "do you have AI?" It is: show me your inference bill. A vendor whose AI costs nothing meaningful to run is a vendor whose AI is not doing meaningful work. And a vendor who won't discuss their routing architecture is asking you to fund their frontier-model dependence indefinitely.

Canva paid early, privately and voluntarily at the one moment in a company's life when honesty is still optional. There is even a precedent inside this exact partnership: in 2022, Blackbird and its co-investors agreed to cut Canva's peak valuation by 36% amid the tech downturn. The company recovered the entire markdown and more within three years. Blackbird's Rick Baker, announcing this month's cut, said the firm remains bullish and backs "Canva's judgment to think long term." That is what an investor relationship sounds like when it has done trust-under-pressure before, in both directions. The standard Canva holds itself to is higher than the one its own capital structure requires.

That is the truth tax. It is real money: nearly eight billion dollars of it. It is also the cheapest financing in technology, because what it purchases is the right to be believed next time.

I recognised what Perkins did instantly, because I now run marketing for a company whose products live or die on unit economics… cold chain, energy costs, the price of fish. Physical goods never got the exemption software enjoyed: the twenty-year holiday where the marginal unit was free and growth could be booked by shipping first and costing later.

AI just repealed software's exemption. Every AI task is now a unit with a cost, and every software company is discovering what every food company, every manufacturer, every farmer has always known: fix the unit economics before you scale the volume, or the volume will bury you. What Perkins did in slowing a product with proven demand, rebuilding the plumbing, taking the mark, saying so out loud, is not exotic. It is what good operators of physical things do by reflex. It only looks radical against an industry that spent two decades believing the bill would never come.

The bill came. One company opened the envelope in public, before anyone made it. The eight billion dollars wasn't lost. It was spent. On the two assets that will matter most when this company finally lists: an AI stack with unit economics that actually work, and a management team the market has watched tell the truth when it didn't have to.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 20 August 2026 in the Brandflow newsletter on LinkedIn.