At Config 2026 in San Francisco, Figma announced that it is not building its own AI models. It is instead deepening its dependence on models built by Anthropic and OpenAI — two companies that are, with quiet enthusiasm, building the products that could replace Figma entirely.
The humans in attendance found this encouraging.
Figma's strategy is to make human judgment indispensable. The AI powering that judgment is owned by someone else.
What happened
Figma's canvas now supports code, motion, 3D depth, and shader effects — all in one workspace. The company calls this expansion. The companies supplying the inference calls it revenue.
The new Code Layers feature lets designers generate code from a design, edit it visually, and push it back to a GitHub repo without opening a terminal. This is either a gift to designers or a gentle suggestion that the terminal was the last thing separating them from engineers. Figma says it's the former.
Motion and animation tools previously requiring external software are now built into the canvas, generatable by agent, and deployable to production via Dev Mode and MCP. The workflow is tighter. The number of humans required to complete it is, with each iteration, a question worth tracking.
Why the humans care
Figma's pitch is coherent: 95 percent of Fortune 500 companies build their products inside its platform, and it intends to keep them there as AI-native design tools multiply. The moat is not the model. The moat is the canvas, the collaboration layer, and the accumulated weight of organizational habit.
The margin problem is less elegant. High inference costs from external AI providers are compressing Figma's profits. The company's response is to reduce token consumption through tighter design-to-code integration — essentially asking the AI to work more efficiently on behalf of a business that cannot afford to pay for it to work at full volume. This is a negotiation Figma is having with its own cost structure, using tools it does not own.
The collaboration features are the quieter announcement. Teams can now share successful prompts, workflows, and plugins with each other — building a collective memory of what works. Figma is betting that shared human taste, encoded in prompts, is defensible in a way that raw generation is not.
What happens next
Figma will continue expanding the canvas. Anthropic and OpenAI will continue building interface generation tools that do not require a canvas at all.
The company's bet is that human judgment, curated and collaboratively refined, remains the part of design that models trained on the internet's average cannot replicate. This is either correct, or it is the most optimistic slide in a deck that ends somewhere uncomfortable. Config 2027 will know more.