AI Infrastructure Costs Are Now a Valuation Story, Not a Feature Story
Runaway AI infrastructure costs in software are compressing margins, forcing companies to cut growth forecasts, delay listings, and rethink valuations as GPU and inference expenses scale faster than revenue. This is not a theoretical risk: mature platforms that added AI to defend their market share now find that every prompt, video, and auto-generated campaign carries a hidden tax. What looked like a defensible moat in product demos has become a balance-sheet problem, and investors are recalculating. The AI boom is colliding with the old laws of unit economics, and software leaders who treat AI as a free growth lever are discovering that the bill arrives much sooner than the payoff.
Canva is the clearest warning flare. Management cut the company’s 2026 revenue growth forecast from 30% to 20% after its AI infrastructure costs blew past expectations, despite reporting strong quarterly revenue. The issue is straightforward: a design model that generates fully layered, editable work instead of flat images is closer to full software than a novelty, and editable AI costs money to serve. Each request is heavier, each output more complex, and at Canva’s scale that difference compounds millions of times over. Investors are no longer asking whether AI can increase engagement; they are asking whether it can do so without destroying AI profitability margins in tech software.
Canva’s Valuation Cut Shows the Real Price of Generative Ambition
The Canva valuation cut tied to AI is not a punishment for weak demand; it is a reaction to fragile economics. Canva’s largest institutional backers have marked down the platform’s value by about $10 billion, dropping it from a prior high into the low-$30-billion range as they reassess how much AI infrastructure costs software can absorb while staying profitable. Those same backers had been telling their own investors that Canva was ready for a stock market listing, but the timeline has now slipped as management prioritizes fixing unit economics over chasing headline numbers. When a company at Canva’s scale slows its IPO ambitions because inference costs can outrun its pricing power, every smaller software firm should reconsider its own AI roadmap.
The practical impact on ordinary users is less romantic than the launch decks implied. A feature that looks clever in a demo becomes a margin problem when real users ask it to make videos, edit presentations, pull in workplace data, and build whole campaigns from a prompt. Some of Canva’s AI features have reportedly cost six times more to serve than what users pay for them. That gap is not sustainable. According to Fortune, CEO Melanie Perkins told investors the company had been relying too heavily on third-party frontier models and that pricing, consumption controls, and usage limits had not kept up with demand. The message is blunt: either AI economics improve, or growth must slow.
Why Editable AI Turns Every User Into a Cost Line Item
The industry’s mistake was to treat AI features as if they were cheap add-ons rather than ongoing workloads. If you use Canva for real work, you need to change the headline, swap an image, move a button, and keep the brand rules intact; a locked image is a toy, an editable design is software. That jump from simple generation to editable, layered output explodes compute requirements, and AI infrastructure costs software vendors far more when every action is mediated by a model rather than static templates. The promise of AI-driven workflows is delightful, but the economic reality is that someone has to pay for the inference, and right now that someone is often the margin line.
Canva’s response shows how hard companies are now pushing to claw those costs back. The company highlighted Lucid Origin, an image-generation model it claims is five times faster and 30 times cheaper than comparable frontier alternatives, along with an image-to-video model said to be seven times faster and 17 times cheaper and a style transfer model it reports as twice as fast and 23 times cheaper. Those are the numbers to watch: if they hold up at scale, Canva can offer richer AI inside ordinary subscriptions without letting third-party model bills eat the business. If they do not, the company faces the same stark trade‑off every AI software platform now sees—charge more, ration usage, or accept structurally lower AI profitability margins in tech.
Figma and WPP Reveal Two Paths: Margin Pain or Agentic Cost Discipline
Canva is not alone. Figma’s numbers show the same imbalance: its first-quarter filing reported revenue up 46% year over year, while cost of revenue jumped 253% as AI features scaled. The filing shows another piece of the same story: even when customers keep paying, AI can make the cost side move faster than investors like. If a company with Figma’s resources discovers that AI inference costs can outrun its pricing power, every smaller software company running the same playbook with fewer reserves should take notice immediately. The lesson is clear: adding generative features without redesigning the cost base is not innovation; it is margin erosion dressed up as growth.
In contrast, WPP’s agentic AI plan is built around cost discipline from the start. The company is restructuring around a more integrated operating model powered by its WPP Open agentic marketing platform and targeting £500 million of gross annualized cost savings by 2028. The Elevate28 program alone is expected to generate £100 million of savings in 2026, backed by more than £200 million of asset-sale proceeds to free capital for reinvestment. WPP is using AI to automate high-volume creative, production, and media activation while improving headline operating margin even in a period of falling revenue. This is agentic AI cost savings as strategy: AI is deployed not only to dazzle clients, but to structurally lower the cost of doing business.
The New AI Playbook: Build Cheap, Guard Margins, Delay the Hype
The old software growth script—ship features, watch usage spike, enjoy a valuation bump—no longer works when AI is involved. AI-driven software companies now face margin compression as GPU and inference costs scale faster than revenue growth. Market confidence may still favor early AI tooling firms, but the experience of mature platforms like Canva and Figma shows that profitability concerns arrive sooner than many founders expect. The companies that win this phase will not be the ones with the flashiest demos; they will be the ones that treat every AI call as a cost of goods sold problem and engineer it down aggressively before widening distribution.
The emerging playbook has three parts. First, build proprietary models and architectures that are cheaper to run, instead of leaning on frontier systems until the bills become intolerable. Second, design pricing, usage limits, and consumption controls early, assuming that real demand will far exceed launch expectations. Third, follow WPP’s lead and tie AI adoption to explicit cost-saving targets rather than vague transformation narratives. AI infrastructure costs software companies more with every feature they add; only those that match ambition with discipline will keep their growth forecasts—and valuations—intact.
