Canva增速腰斩：AI成本、代理决策与估值重估

The week the “AI is just a feature we’ll add” argument stopped working, and three pay bands replaced two.

This week on 20VC x SaaStr with Harry Stebbings, Jason Lemkin and Rory O’Driscoll, we started somewhere other than OpenAI and Anthropic for once: Canva cutting its 2026 growth forecast by a third as AI serving costs blow up. That led into Google losing Jeff Dean and Demis Hassabis, Elon’s $16.8B first installment on Terrafab, Revolut’s reported $50B CEO package, and Whatnot raising at $20B.

1. Canva going from 30% to 20% growth is an existential question, not a gross margin one

Canva did roughly $3B in GAAP revenue last year and entered 2026 growing 30%. Melanie Perkins disclosed mid-year that they’ll likely finish the year around 20%. Part of the stated reason: the AI features they’re shipping cost real money, and subsidizing frontier model calls across a prosumer base is expensive.

Rory’s read: the margin math is a second-order business model issue. What matters is whether 30 to 20 is on the way to 10. There are three large creative software companies right now: Adobe at $23B growing 12% and trading at three to four times revenue, Figma public at $1.4B growing 40%, and Canva private at ~$3.6B growing 20%. All three face the same question. Is AI a feature you incorporate, or a new thing that makes you obsolete?

2. Our agents never once suggested Canva

We churned Canva. We also churned Notion. Neither did anything wrong. We just no longer had a need for them in an agentic operating model.

Jason’s read: we built our own ad server and creative generation network at SaaStr, running entirely on our agents. It never occurred to the agent to use Canva for any of it. Not once. Users switching to a competitor is a fixable problem. The layer making the decision not having your product in its consideration set is a different kind of problem.

Harry connected it to what he heard from the President of Uber, whose single biggest fear is the disaggregation of UI: you say you want a car, and ChatGPT routes you to whichever provider is cheapest. Our agents already did that to Canva.

3. The era of no-code is winding down, and it covers more companies than people count

Amjad at Replit said it about Airtable, not Canva, but it applies to all of them. Airtable was a no-code database disguised as a spreadsheet. Notion is a no-code database disguised as a word processor. Canva was a no-code way to design things, and it was breathtakingly disruptive: you didn’t need a designer, you didn’t need HTML.

Jason’s read: the entire category of “things humans can do without engineering resources” is winding down, because AI does those things natively. If the job can be done inside ChatGPT or Claude, even accidentally, the standalone tool is in trouble.

4. Figma is more protected than Canva, and the reason is bureaucracy

Rory’s read: Figma is an enterprise product. Large groups of people coordinating on building software. Even if you fully automate the creative act, you still have corporate process and workflow to monetize. Canva is the prosumer motion: I want a flyer, I want a cheap website, I want a piece of content. That’s where AI is most accessible, because you type the request and get the output.

Figma took its hit publicly. Dylan was clear that gross margins will be meaningfully impaired as agentic products get used, and the stock traded down 20%. Taking that hit is more painful in public than in private, which is part of why Canva’s situation reads differently even with worse numbers.

5. Canva is probably worth around $12B today

Jason’s read: 20% growth at ~$4B ARR against current public comps, decelerating rather than accelerating. There’s some rule of 40 adjustment, but $12B is where it lands.

Rory’s pushback: you can find companies at 20-25% GAAP revenue growth, free cash flow positive, trading well above that. Datadog, Cloudflare, JFrog: mid-20s growth, 20%+ operating margins, 15 to 17 times NTM. The difference is that nobody is asking whether those companies exist in five years. If Canva carries the existential discount, $12B is right. If they transcend it, it’s $12B and up. Either way, liquidity only comes at the end of the journey, and that journey is at least 12 months long.

Harry’s note for LPs holding Canva: when the private window shuts and appetite dies, getting anything done at scale becomes very difficult.

6. After Airtable and Canva, be extra skeptical of marks

Jason’s read: these are two events that quietly hit old marks and are difficult to hide behind. You should probably have marked them down last year. “We raised at $42B, therefore we’re holding at $42B” is no longer defensible.

Rory’s method: ground yourself in the actual growth rate first. We’re doing a billion growing at 30%, that’s this multiple. We’re doing a billion growing at 10%, that’s a very different multiple. Only then do you grade up or down for existential risk versus AI lift.

7. Selling along the way is mathematically true and strategically dangerous for small funds

Freestyle’s Dave Samuels noted their blended exit price on Airtable was around $6B, making the case for trimming in good times.

Rory’s data: roughly 70% of the time you should have sold, 30% you should have held. But the holders compound forever, and the ones you sold compound at zero from that point on. It’s the Bessembinder result: sub-1% of companies deliver 90% of the capital gains in public markets, and it’s the same in private.

Jason’s read: for a smaller fund, the math gets tricky fast. I’m lucky to have three fund returners. If I start taking early exits on those and never get a 10x fund returner, the LPs don’t get the 5x, 6x, 8x fund they’re underwriting. I ran the analysis across my entire history, angel and venture, on who I should have sold and who I should have held, and it broke roughly 50/50.

8. Companies that hesitated in 2023 and 2024 are paying the bill in 2026 and 2027

Rory’s frame: there are moments when a crevice opens between the before and the after, and if you’re going to make the jump you have to make it fast. Eventually the gap is too wide.

The best example on the apps side is Palantir, which went from 18% growth to 98% growth. Two things made it work. First, real forward deployed engineers, people who’d spent a decade and a half deploying operational change in the field, so when customers needed AI they already had the humans to do it. Most companies have solution architects wearing an FDE t-shirt. Second, genuine outcome-based pricing. Give me $2B, I’ll save you $8B. Everyone talks about outcome-based pricing. Almost nobody puts a $2B contract on the line for it.

Jason’s read: Replit is the harder version of the same story. Six years in the wilderness as a nerdy web IDE, then the models arrived and they rethought everything. Palantir could have chugged along at 20% and been fine. They grabbed the moment instead.

Worth separating out: Datadog, Cloudflare, Palo Alto and Twilio didn’t have to reinvent anything. They sell infrastructure into the largest infrastructure boom in history. Showing up and selling more is a very different job than the one facing Canva’s CEO.

9. Jeff Dean left over compute allocation, not money

Jeff Dean left Google after 27 years, taking three researchers with him. Demis moved into a chairman role. The stock dropped a couple hundred billion in market cap, which as Rory noted is an unusually validating exit for a non-CEO.

Rory’s read: look at how compute gets allocated at Google. Compute given to Cloud converts into 30% operating margins immediately, because they can sell it to Anthropic. Compute given to Gemini might produce a competitive coding model and consumer revenue. Compute given to drug discovery or materials science is a five to seven year moonshot indulged at the 10% level. If your work is permanently third in line, you leave.