- The case: Anthropic and Figma cooperated in early 2026 (“Code to Canvas”, February). On 17 April Anthropic launched Claude Design, a competing product — Figma’s stock fell 6–7.7% per reports.
- Platform risk has two faces: availability (the model gets pricier, deprecated, switched off — see the 19-day Fable 5 shutdown) and competition (the provider rebuilds your product).
- The answer is not “no AI” but sovereignty through architecture: code ownership, swappable models, open models for sensitive data.
- Even those who — like us — use frontier models stay independent as long as the model sits behind an abstraction and the code belongs to the client.
Some stories from the tech world sound like distant noise to the mid-market — and then, on closer look, describe your own situation. The Figma/Anthropic case is one of them. At first glance a power struggle between two US firms over the design market. At second glance a blueprint for what happens when you leave the foundation of your business to a provider that can move up the value chain itself at any time.
What exactly happened
The chronicle is short and clear. In early 2026 Figma and Anthropic collaborated: in February they built “Code to Canvas”, which turned Claude-generated code into Figma designs — a classic partnership in which Anthropic’s model enhanced Figma’s product. Two months later, on 17 April 2026, Anthropic launched Claude Design: a tool that generates prototypes, presentations and marketing material directly from prompts — running on Claude Opus 4.7. With it, Anthropic pushed into exactly the market that historically belonged to Figma, Adobe and Canva. The market’s reaction was unambiguous: Figma’s stock fell by 6 to 7.7% according to reporting.
Important for context: this is no breach of contract and nothing illegal. It is normal platform behaviour. Whoever owns the model infrastructure sees which applications succeed on it — and has every incentive to build the most lucrative ones itself. That is exactly what makes it a structural risk for everyone who builds on that infrastructure, not a one-off.
Platform risk has two faces
The Figma case shows one side. The other we saw a few weeks ago elsewhere. Together they form the full picture:
| Risk | What happens | 2026 example |
|---|---|---|
| Availability | The model gets more expensive, deprecated or switched off | Fable 5: 19 days offline worldwide by US order |
| Competition | The provider learns from the integration and rebuilds your product | Claude Design vs. Figma, two months after the collaboration |
| Rulebook | Terms, data protection or access change unilaterally | Subscription plans shifting models to usage credits |
Framing: Digital Maker — the three forms of platform risk
We described the availability side in detail in the chronicle of the Fable 5 shutdown: a frontier model was gone for every customer for 19 days, for reasons no customer could influence. The Figma case adds the competition side. Both lead to the same conclusion: an AI provider is not a tool you buy, but an ongoing dependency you manage.
Why this hits the mid-market in particular
Large corporations have legal departments, fallback contracts and bargaining power. An owner-led business does not. If a mid-sized company hard-wires its proposal generation, its customer communication or an own digital product onto a provider ecosystem — its connectors, proprietary plugins, specific subscription features — then a piece of operational capability hangs on decisions made in California. And unlike cloud storage, which you can migrate if you must, a deeply integrated AI solution has half its business logic stuck inside the ecosystem.
This is not scaremongering and no reason to avoid AI — on the contrary. It is a reason to introduce AI properly. The difference between “dependent” and “sovereign” lies not in whether you use powerful models, but in how you integrate them.
Sovereignty through architecture — the four levers
- 1. Code ownership. The solution, its logic and its configuration belong to you, live in your accounts and are documented. The provider does not own your process — you do.
- 2. Abstraction layer. The model is not wired directly into every process but addressed through a thin layer of your own. A provider switch becomes a configuration change — not a project. (Why this is the decisive building block is in Which AI model does the mid-market need?)
- 3. Data sovereignty. Sensitive data runs on self-operated, open models (guide in German) — locally or in a European cloud. What never leaves the building cannot be mined, repriced or used against you by any provider. The strongest open models in 2026 increasingly come from China (open-weight), plus European options — an overview is in ChatGPT alternatives from Europe.
- 4. No core process on a pure provider ecosystem. Connectors and proprietary plugins are handy for the non-critical. For business-critical flows, control belongs in-house — open standards instead of ecosystem lock-in.
“But you use Claude yourselves” — the honest part
This question has to be allowed when you argue this way, and we ask it of ourselves. Yes, Digital Maker builds on the Claude API — for many demanding tasks it is simply the best tool. The point is not to avoid frontier models. The point is not to chain yourself to them. Because we build code the client owns and the model sits behind an abstraction, every provider is swappable for us: were Anthropic to become a client’s competitor tomorrow, double the price or go offline for 19 days again, we swap the model and keep working. That very way of building separates a bought dependency from a used tool — and it is why the Figma lesson is a confirmation for our clients, not a problem. What to look for when choosing an implementation partner is in Choosing an AI agency — the question “Who owns the solution in the end?” is not there by accident.
What the mid-market should do now
- Take stock: which of your processes hang on which provider ecosystem — and how deeply? What could be migrated in days, what is stuck?
- Separate the critical from the non-critical: business-critical flows need ownership and swappability; for edge functions, ecosystem convenience is perfectly fine.
- Sovereignty by data sensitivity: sensitive data on open, self-operated models; non-critical volume wherever it runs cheapest and best.
- On new projects, ask from day one: do we own the code afterwards? Is the model swappable? If not — renegotiate or build differently.
The Figma lesson fits in one sentence: the platform that makes you strong today can replace you tomorrow — if you leave it in control. The good news for the mid-market is that the alternative has long been practical. Whether and when self-operation pays off against buying, we work through in the build-vs-buy guide (in German).
Sources and context
The account of the Figma/Anthropic case is based on public reporting (incl. VentureBeat, The New Stack, Fast Company, Cryptobriefing) on the launch of Claude Design on 17 April 2026 (running on Claude Opus 4.7), the prior collaboration on “Code to Canvas” in February 2026, and the reaction of Figma’s share price (a fall of roughly 6–7.7%), as of early July 2026. The details on the Fable 5 shutdown follow Anthropic’s public statements. Assessments, the risk taxonomy and the architecture recommendations are Digital Maker’s view based on our project experience and are not legal or investment advice.
FAQ: platform risk with AI
What happened in the Figma–Anthropic case?
Figma and Anthropic collaborated in early 2026 — in February 2026 on a feature called “Code to Canvas” that brought Claude-generated code into Figma designs. Two months later, on 17 April 2026, Anthropic launched “Claude Design”, a tool that turns prompts into prototypes, decks and marketing material — aiming straight at Figma’s market. Figma’s stock fell by 6 to 7.7% according to reporting. The partner had become a competitor.
What does platform risk mean with AI?
Platform risk means: you build your business or a product on a provider’s infrastructure — and that provider can change the rules, restrict access or enter your market itself. With AI it has two faces: availability risk (the model gets more expensive, deprecated or switched off) and competition risk (the provider learns from the integration and rebuilds your product). Both became real in 2026.
Are local or open-weight models the answer to platform risk?
For a growing share of workloads, yes. An open-weight model you run yourself — locally or in a European cloud — cannot be switched off, repriced or pushed out of the market overnight by anyone. You control version, location and access. The price is operational effort and often some top-end performance. Hence the realistic answer is hybrid: frontier models where they are irreplaceable, open models for the sensitive, high-volume mass.
Isn’t Digital Maker itself dependent through its Claude usage?
A fair question — and the answer is architecture. We build code the client owns, and we address the model behind an abstraction layer. If a provider fails, becomes too expensive or turns into a competitor, the switch is a configuration change, not a rebuild. Dependency doesn’t come from using a model — it comes from binding yourself inextricably to an ecosystem.
How does a mid-sized company protect itself from platform dependency?
Four levers: first, code ownership — the solution and its logic belong to you, not the provider. Second, an abstraction layer that keeps the model swappable. Third, data sovereignty — sensitive data runs on self-operated, open models. Fourth: never build business-critical processes on pure provider ecosystems (connectors, proprietary plugins) whose rules can change at any time.
How dependent is your business on a single AI provider?
In a discovery call we take stock: which processes hang on which ecosystem, what can be shifted to your own models for sensitive data, and where an abstraction layer makes you independent. Four eyes, thirty minutes, no slides.