In May 2026, Arthur Mensch stood before the deputies at the National Assembly, and he said something I find quite dizzying when you really think about it. He spoke of an "inability of players to disengage from these new services." Not a difficulty, not a brake, not a potential risk... an inability. Meaning the companies that have switched to the APIs of OpenAI, Anthropic, or Google, well they can no longer leave. Or at least at a price they haven't even begun to calculate, and that's what concerns me the most.

"an inability of players to disengage from these new services."

When I read that, it reminds me of something. It reminds me of Microsoft in the 2000s, when Office, Exchange, and Active Directory came in through the door of productivity and stayed through the door of dependency... and back then, leaving Microsoft was no longer a technical choice, it was a multi-year project, a complete migration plan, a rewriting of all processes, a training of the entire organization. Many companies never finished. Others never even started. Well, AI is the same movie, except this time it's worse, and I'm going to explain why.

Your data served to contextualize their models

The first difference from the Microsoft lock-in is that when you were on Office, your data was in a closed format but at least it was readable. With AI, your data served to contextualize a model that doesn't belong to you, and that's fundamentally different. You built vectorization bases for RAG, you did fine-tuning on your internal documents, you wrote system prompts of several thousand characters, tested, optimized, versioned... well all that work is tied to the specific API of the provider you chose.

Prompt formats are not portable from one provider to another. The function calling schema differs between OpenAI and Anthropic. The vectors computed by an OpenAI embedding are not compatible with those of a Mistral model. You've invested hundreds of hours of engineering to calibrate a system on one provider, and this capital doesn't travel, it's nailed down. The lock is no longer on the file like with Office, it's on the entire inference chain, and that changes everything.

Your workflows depend on the specific capabilities of the API

The second difference is that modern workflows don't just send text and receive text. They use function calling so the model triggers actions, they rely on structured output to get valid JSON, they send images for vision, they chain calls, manage context windows, do prompt caching... and all these capabilities exist at the main providers, but with different syntaxes, different behaviors, different limits.

OpenAI's function calling doesn't behave like Anthropic's. Google's structured output doesn't have the same constraints as Mistral's. Switching models can break pipelines, and not cleanly, not gradually... in production, in the middle of the night, on an edge case nobody had tested. I see this with the clients who come to see me... they have an agent running on an American provider's API, it works, it brings in revenue, but they no longer dare touch the configuration because the slightest model change triggers a cascade of regressions to retest by hand. The system has become an implicit monolith, and everyone pretends it's normal. It's not normal.

The cost of migration grows over time

The third difference is perhaps the most insidious... the cost of migration grows over time. The longer we wait, the deeper the integrations, the more teams are trained on one tool and know only it, the more internal documents and runbooks are written for a specific API, the more client contracts embed guarantees tied to a specific model... it's a debt that accumulates in silence, that appears in no financial statement, no dashboard, no project review.

But the day the provider changes its pricing, cuts access on a government order like Anthropic did in June 2026, modifies the model's behavior in a new version... that day, the debt becomes payable. And it's paid dearly. When your AI provider obeys a government other than yours, what guarantees that your service will still run tomorrow? A contract... a contract the other party can suspend.

Open source, the only real safeguard

So yes, there's only one model that truly protects against lock-in, and I'm not saying that out of ideology, I'm saying it because it's a verifiable technical fact... open source. The model is yours, you can run it wherever you want, for as long as you want. Nobody can cut access. Nobody can change the pricing, since you pay for electricity and the machine, not an inference token. Nobody can modify the model's behavior without you deciding it. It's as simple as that.

Arthur Mensch, still before the deputies:

"In a world where open source wins, prices are much lower because the margin is not made on the models."

The margin isn't made on the model, it's made on the service, the integration, the support. That's exactly what I do with my clients... not an inference token to pay, but the setup of a local model, team training, RAG configuration on their documents. The model itself is free.

Lock-in is being set up right now

Lock-in is a process that's being put in place right now, silently, with every API integration, every workflow built around a specific capability, every team training on a proprietary tool, every contract tying a service guarantee to a specific model. Each year of delay increases the cost of exit. Each additional integration adds another bar to the cage.

After three years, leaving is no longer a migration project, it's an enterprise project, with its budget, its steering committee, and its non-zero probability of failure. The question isn't whether you should leave OpenAI or Anthropic... the question is how much the departure will cost in two years, and who will pay. Because there will be a departure, sooner or later, the provider will change its terms, a government will cut access, the price will become untenable. And that day, the cost of migration will be the total of all the integrations we agreed to write on an API we didn't control.

You're writing the rest of this scenario today, with every API call. So be very careful about what you write tomorrow.