Canva has 260 million monthly active users, 28 million paying customers, $3.5 billion in annual revenue, and 95 percent of the Fortune 500 as customers. Its biggest problem right now? Users are too active.
In April 2026, Canva shipped AI 2.0: type one sentence, get a fully editable design, document, website, or marketing asset. AI moved to the center of the product. Four months later, the company quietly slowed its AI push and cut its 2026 revenue growth guidance from roughly 30 percent to about 20 percent. The reason is not demand. It is the bill. Every AI generation consumes GPU and inference resources, and Canva's free tier still hands out 200 standard AI calls or 20 premium calls per month to users who pay nothing.
Canva is not an outlier. Figma reported Q2 revenue up 48 percent year over year, while cost of revenue jumped 117 percent, from $27.9 million to $60.5 million. GAAP operating loss reached $117.3 million, against an operating profit of about $2 million a year earlier. The stock fell 15 percent after hours.
Software used to be the lightest business on earth: every extra user cost almost nothing. AI is making it heavy. Consumer software — one of the best businesses of the past 20 years — is quietly turning into a gym. It wants you hooked, and it is terrified that you will actually use it.
Why AI Broke the SaaS Flywheel
Before AI, Canva ran the textbook SaaS playbook. A free online editor pulled in massive traffic; premium templates and licensed assets converted users into Pro and enterprise plans. Every feature had the same property: the more people used it, the lower the unit cost. Scale and profit moved in the same direction.
AI inverts that equation. Opening a template costs Canva almost nothing. Generating an image or calling a model costs real money every single time. For the first time, the free traffic that powered growth shows up on the income statement as a cost.
The second problem is pricing power. In 2024, Canva tried to migrate early Teams customers to a per-seat model: some US users had paid $120 a year for five seats; the new standard was $10 per user per month with a three-seat minimum — roughly $600 a year for the same team, a four- to fivefold increase. Longtime users revolted, and Canva was forced to grandfather early Teams plans at historical prices. Canva won its users from Adobe by being simple and cheap. It cannot raise prices without betraying the identity that made it.
Three Pricing Models, Three Failure Modes
Since AI pushed costs up, the industry has tried exactly three pricing models — and each one breaks in a different way.
Pure subscription: revenue is capped, compute is not. Adobe raised the flagship plan from $52.99 to $59.99 after Firefly launched in 2023, then introduced Creative Cloud Pro at $69.99 in 2025. The latest quarter still shows the structural gap: subscription revenue up 13.7 percent, subscription cost up 16 percent. A subscription fixes income; AI usage keeps floating.
Usage-based pricing: risk moves to the user, and bills become unpredictable. Replit adopted "Effort-Based Pricing" in 2025, charging by the workload an agent completes — roughly $0.25 for a simple task, over $1 for complex ones. Heavy users now pay for the compute they burn, but an agent plans, searches, calls tools, and retries in the background; the user sees one task, not the token bill. In July 2025, a billing calculation error overcharged about 6 percent of Replit's paid users, forcing refunds plus $10 credits. Developers reported projects that used to cost $2–3 now pricing at $30, and one user posted a farewell after a single failed task deducted $32. The deeper problem: once everything is billed like model calls, users start asking why they should not just buy the API directly.
Outcome-based pricing: the fairest in theory, the hardest in practice. Zendesk has charged roughly $1.50 per AI-automated resolution since 2024 — you pay only when AI actually solves a ticket. The catch is defining "solved." Under Zendesk's early rules, a conversation counted as auto-resolved if the user did not reply within 72 hours. One customer audited 192 auto-resolutions and estimated that at least 40 percent still needed human review. It is the same verification gap that haunts AI content provenance, where SynthID-style text watermarks are being broken almost as fast as they ship. Tokens are easy to count. Outcomes are not.
The Industry Answer: Hybrid Pricing
After cycling through all three, most AI applications are converging on the same answer: hybrid pricing. Collect a base subscription, then charge separately for excess AI usage or for outcomes.
Intercom sells a $39-per-month seat and charges $0.99 per resolution. Figma introduced AI Credits in March 2026 to meter usage. In April, both Figma and Canva began splitting fixed subscriptions into "base plan plus overage."
Look past the increasingly complicated price lists and the real change is visible: the industry is reallocating compute risk. Under pure subscription, the company bears it — the more users consume, the higher the cost. Under pure usage pricing, the user bears it — a task's price becomes impossible to predict. Outcome pricing just relocates the dispute to the definition of "done." Hybrid is the pragmatic equilibrium: subscription revenue protects the floor, and overage pushes the cost of power users back out.
Deeper Than Pricing: App-Layer Pricing Power Is Eroding
The pricing patch does not hide the deeper shift. Once every feature is billed like raw compute, users start doing a different calculation: if it is all token calls anyway, why not buy the OpenAI or Anthropic API directly?
That is the question hanging over Salesforce's Agentforce, whose annual recurring revenue passed $1.2 billion, up 205 percent year over year — while the market keeps asking whether customers are paying for Salesforce's accumulated data, workflows, and enterprise software, or for a thin wrapper around model calls. The same logic that turned Oracle's database into an agent platform applies here: the companies that keep pricing power are the ones whose software carries proprietary data and workflows, not the ones reselling tokens. We covered Oracle's agent-platform bet earlier this week.
Two consequences follow. Free users — the growth engine of the last two decades — have flipped from an asset into a liability on the income statement. And the software premium itself shrinks: the convenience of packaging is worth less when customers can compare your per-call price against a raw API price.
What This Means for AI Builders
None of this means the consumer AI business is dead. It means the playbook changed:
- Price your compute exposure on day one. Unlimited free tiers are a growth strategy for zero-marginal-cost software; for AI they are a burn strategy. Meter aggressively from the start.
- Define the outcome unit early. Zendesk's "what counts as solved" fight shows that whoever defines the unit of value controls the pricing conversation. Do it before your users do.
- Defend app-layer value. Proprietary data, workflows, distribution, and switching costs are what keep your price above the API price. Without them you are a wrapper, and the market will price you like one.
- Model AI cost as COGS. Track cost per account and usage per account as core metrics, not footnotes. Margin compression is now a design constraint, not an accounting surprise.
The gym-membership era of software is not a failure of AI — it is the economics of AI showing up in the income statement. The winners will be the companies that treat inference cost as product design, not as a problem for finance to solve later.
