On April 14, Mike Krieger, Anthropic’s chief product officer, resigned from Figma’s board of directors. The same day, it was reported that Anthropic’s next model, Opus 4.7, would ship design tools that compete directly with Figma’s core offering. Three days later, Anthropic launched Claude Design.

That sequence is worth looking at closely, not for what it says about Figma or Anthropic specifically, but for what it says about any company building its operation on a single AI vendor’s layer.

The partner that becomes a competitor in three days

We’ve written before that the moat has moved from the model to the harness: the workflow layer wrapped around the model that decides how it actually gets used. Figma spent over a decade building that harness for design. Anthropic built the model Figma used to power its own AI feature, Figma Make. That was the relationship: Anthropic supplied the intelligence, Figma supplied the workflow.

The problem with that relationship is it only works as long as both sides agree on where each layer ends. On April 14, Anthropic decided it no longer agreed. Krieger left the board the same day news of the competing model broke, and three days later Claude Design was live in production: an AI workspace that doesn’t just assist designers, it aims to remove the need for one during the early build phase of a product.

Figma’s stock dropped nearly 7% the day of the launch. CEO Dylan Field later said Anthropic was “not consistently candid in their communications,” a phrase he didn’t choose by accident. It’s nearly word for word what OpenAI’s board said when it briefly fired Sam Altman. Field knew exactly what he was quoting.

Figma’s defensive move

This week, three months after that launch, Figma acquired the team behind Bud, formerly known as Orchids, a Y Combinator-backed vibe-coding app. Bud and Orchids shut down on July 18; the team folds into Figma to build its own code-generation capability.

The obvious read is that Figma is now covering both flanks at once. From above, Anthropic built Claude Code into a $2.5 billion annualized-revenue machine, with 71% of developers who regularly use AI agents naming it their primary tool, according to a Pragmatic Engineer survey of 15,000 developers. From below, startups like Lovable (reportedly in talks to double its valuation to $13.2 billion this week) and Replit (valued at $9 billion, tripling in six months) are building the direct prompt-to-product path with no traditional design tool in between.

Buying the Bud team is Figma’s answer to that squeeze: if the AI lab lending you the model can decide one day to compete with you in your own layer, the only real defense is to stop depending on a single provider for the piece you’re missing.

Bud didn’t arrive spotless, either. The BBC had previously reported, citing a security researcher, that apps built on Orchids (Bud’s earlier name) were vulnerable to attacks. Figma didn’t buy a finished, polished product. It bought a team and a technical base that still needs real security work before enterprise customers should trust it in production. This is a speed play, not a guaranteed-quality one, and that distinction matters if you’re a Figma customer wondering how fast you’ll actually be able to use this new capability for real work.

Why this isn’t just Figma’s problem

The temptation is to read this as a Silicon Valley anecdote between two companies that don’t touch you directly. The pattern underneath is broader, and it does touch you if your company is building anything important on top of one AI vendor’s API or deep integration: we already saw the same dynamic with Google Stitch entering the same design-to-code territory, and we’re going to keep seeing it because the line between “model” and “product” got blurry the moment AI labs realized they can capture more margin building the application layer themselves instead of selling it to you as infrastructure.

No partnership agreement protects against this. Krieger had a board seat at Figma. That stopped nothing. The only real protection is architectural: build your operation so you can swap model providers without everything you built on top collapsing, and keep early visibility into which direction each lab is heading before they announce it.

In practice, that architectural protection comes down to a few concrete habits. Keep the layer where your real competitive advantage lives (your workflow, your data, your customer relationships) separate from the layer where you’re just consuming a model’s intelligence, so a change on one side doesn’t drag down the other. Don’t integrate so deeply with a single provider that switching models turns into a full rewrite instead of a configuration change. And treat any “strategic partnership” with an AI lab for what it actually is: a commercial relationship between two companies competing for the same territory of value over the medium term, not a permanent alliance.

At IQ Source, that’s literally the first question we answer in AI Maestro, our two-month discovery program: how dependent does your operation become on a single vendor once we design the architecture, and what happens if that vendor changes strategy tomorrow. It’s not a hypothetical question. It happened to a $10 billion public company with a board seat inside its own most important vendor. A mid-market company without that level of access can have it happen with a lot less warning.

The question I’d ask any technology leader today isn’t how good your AI vendor is. It’s how exposed you are the day they decide the layer where you create value is also a market they want to own.