When your whole market picks the same vendor

Most advice about vendor lock-in stops at the export button. Check that you can download your data, keep your prompts in a text file you own, avoid the annual contract, and you are told you are covered. That advice is correct as far as it goes, and it goes about half the distance. The other half is not in your account settings at all. It is in what everyone around you has already decided to use.

The half you control is the work sitting inside one company’s product. The half you do not control is the version where your clients, your suppliers, the platforms you sell through and the tools you buy have all converged on the same vendor, so that choosing differently costs you customers rather than an afternoon. No export button touches that. This guide is about telling the two apart, pricing the second one honestly, and deciding what is worth doing about it. It is written for a solo operator or a team under about twenty spending its own money. If you run procurement at a company large enough to negotiate bespoke contract terms, you have levers this guide does not assume you have.

Two kinds of lock-in, and only one has an export button

Split your stack into two lists, because the two respond to completely different remedies.

The first list is everything holding material you made: documents, tables, prompts, automations, client records. This is chosen lock-in. You bought it, you filled it, and in principle you can empty it. The remedy is an export, and the honest thing to say about exports is that they work and they are lossy. Notion will return your workspace as Markdown and CSV, HTML, or PDF, which is more than many tools offer. A workspace export can also take up to 30 hours to process depending on its size, the download link expires after 7 days, and a Form view of a database cannot be exported at all [8]. So the export exists, it is slow, and it hands back your content without some of the structure that made the content useful. That is the normal case across the category, not a complaint about one product.

The second list is everything you use because somebody else decided. The file format a client sends. The assistant your biggest customer’s team already lives in, so your deliverable has to arrive in a shape that opens there. The connector standard your tools happen to speak. The platform that owns the audience you sell to. None of that is yours, so none of it appears in an export, and none of it improves because you were careful when you signed up. Call it market lock-in. It is the more expensive of the two and it gets almost none of the attention.

Switching cost compounds through other people’s decisions

The clearest recent illustration is not a small business, which is exactly why it is useful. On 15 July 2026 Nvidia announced Cosmos 3 Edge, a 4-billion-parameter model “lightweight enough to run on edge GPUs” that predicts robot actions locally rather than in a data centre, and published alongside it a roll-call of “Japan’s physical AI ecosystem leaders” that “intend to join the coalition”, among them FANUC, Fujitsu, Hitachi, Honda R&D, Kawasaki Heavy Industries, Kubota, Mitsubishi Corp., NEC, Preferred Networks, SoftBank Corp., Sony Group Corporation, Telexistence and Yaskawa Electric [1]. The same day, Nvidia’s Japan write-up reported that Mizuho plans to build what is expected to be the largest on-premises AI factory in Japan’s financial industry, starting with DGX B200 systems, that RIKEN is deploying two systems using 1,600 and 540 Blackwell GPUs, and that Japan had launched a government-backed physical AI initiative [2].

Taken one at a time, those are ordinary procurement decisions, each defensible on its own merits by the company making it. Taken together they are something else. Once your suppliers, your competitors, your national research institute and the reference systems your regulators see all speak the same stack, the cost of choosing differently stops being your engineering cost. It becomes the cost of being the one party in your industry that cannot exchange work with everybody else. That cost rises every time somebody else joins, which means it rises without anyone deciding to raise it.

Note also how the commitment is framed. Sovereignty is the selling point. Hardware sits on the premises so that a bank can build agents “while keeping critical data close and secure”; one partner is described as building “sovereign financial intelligence” on Blackwell GPUs; the government initiative is aimed at open multimodal foundation models for agents, digital twins and robotics [2]. Both things can be true at once. You can hold the machines and hold the weights and still be standing on one foreign company’s architecture. Sovereignty over the copy is not independence from the design.

You are not Mizuho, and the numbers above are not your numbers. The shape still applies at your size, and it is easier to see there. If every client in your niche now sends briefs as a shared doc inside one vendor’s workspace, your preference for a different tool is a tax you pay on every job, not a principled stand anyone is going to reward.

Depth of commitment predicts the cost of a move better than the size of the bill

The instinct is to rank exposure by spend. The vendor taking $200 a month feels riskier than the one taking $20. Spend is a poor predictor. What predicts the cost of a move is how many other people would have to agree to it.

At the far end of that scale sit commitments that have been certified. A driving stack validated against one vendor’s software does not get re-pointed at a competitor with a configuration change, because the validation, not the software, is the asset that took the time. Nvidia’s Japan write-up shows the pattern in one line: Toyota is building next-generation vehicles on NVIDIA DRIVE AGX “running the safety-certified NVIDIA DriveOS operating system”, and using a custom automotive model to generate and validate safety-critical code “while adhering to stringent automotive compliance” [2]. A commitment that sits inside a compliance argument is the hardest kind to reverse, because reversing it means making the argument again.

Your version of certification is smaller and behaves identically. It is anything you have written into a client contract, put in a signed proposal, trained a contractor to use, built a documented process around, or named in an answer to a client’s data-protection questionnaire. A tool you use alone and undocumented is a one-hour switch. A tool named in six client contracts is a renegotiation with six counterparties who have no reason to help. Rank your stack by who else would have to say yes, and the order changes immediately, usually in a way that has nothing to do with the invoice.

Compatibility layers buy you an afternoon, not a strategy

Vendors know portability sells, so several of them now accept requests written for a competitor’s library. Anthropic publishes an OpenAI SDK compatibility layer: you point the base URL at https://api.anthropic.com/v1/, swap in a Claude API key, and change the model name [5]. The Gemini API offers the same trick, with a base URL of https://generativelanguage.googleapis.com/v1beta/openai/ [6].

Then read the caveats, because both vendors wrote them down. Anthropic states that the layer is primarily intended to test and compare model capabilities and “is not considered a long-term or production-ready solution for most use cases”, and directs you to the native API for PDF processing, citations, thinking and prompt caching [5]. Google states that support for the OpenAI libraries “is still in beta while we extend feature support”, and, on its image endpoint, that parameters outside the documented list “will be silently ignored by the compatibility layer” [6].

Silently is the word to sit with. A compatibility layer that raises an error on an unsupported option is a portability tool. One that drops the option and answers anyway is a way to run a fortnight of quietly degraded work without noticing. Use these layers for what their own documentation says they are, which is a fast way to run your real prompts against a second model before you commit to it. Do not file the existence of the layer as a migration plan.

Prefer the standard that nobody owns

The durable defence against market lock-in is not backing the underdog. It is preferring the parts of your stack governed by something other than the company selling them to you.

The Model Context Protocol is the current example worth copying. It is an open-source standard for connecting AI applications to external systems, supported by Claude, ChatGPT, Visual Studio Code and Cursor among others, which is what lets you “build once and integrate everywhere” instead of writing the same connector twice [3]. More to the point, it has been established as “Model Context Protocol a Series of LF Projects, LLC”, code and specification contributions are made under the Apache License 2.0, and the governance document states that membership in the technical governance process is for individuals rather than companies, with no seats reserved for specific companies [4].

That is the test to apply to anything described as open, and the answers are usually published, so the check costs about ten minutes. Who holds the trademark. Which licence the specification itself is published under. Who is allowed to change the specification, and whether that group can be reshaped by one company’s staffing decisions. Whether an implementation still runs with the originating vendor switched off. A format that fails those four is a proprietary format wearing a friendly word.

The law gives you one lever, in one place

If you are in the EU, or buying from providers who serve it, there is now a statutory floor under cloud portability that did not exist a few years ago. The Data Act has applied since 12 September 2025 and sets a framework for customers to switch between providers of data processing services [7]. On money it is staged. As a transitional measure “from 11 January 2024 to 12 January 2027” providers “may still charge their customers for the costs incurred in relation to switching and data egress”, and from 12 January 2027 the Act will “entirely remove switching charges, including charges for data egress” [7].

Be precise about what that buys. It bites on data processing services, so it reaches the hosting and platform layer where your files and workloads sit. It does nothing at all about the fact that your three largest clients work inside one assistant, because that is a preference of theirs, not a contract you are a party to. The practical use is as a question you ask before signing: what does your switching process look like after January 2027, in writing. A provider that cannot describe it in a paragraph has told you something worth knowing.

Measure concentration, not lock-in

Lock-in is a yes or no, and a yes or no flatters whoever is asking. Concentration is a number, and a number can be checked once a quarter without a meeting.

The number worth tracking is the share of your revenue you could not deliver next month if one vendor’s ecosystem became unavailable, unaffordable, or unacceptable to a client. Not the share you use, the share you could not replace. Work you do inside an assistant and hand over as a finished document is close to zero exposure, because the assistant is interchangeable and the client never sees which one it was. Work that exists only because you are listed, integrated, certified or visibly fluent in one vendor’s platform is close to total exposure, because losing the platform loses the work.

Then decide, on purpose, whether that number is one you want. High concentration is not automatically a mistake. Being the person in your market who is genuinely excellent with one platform is a real business, and spreading yourself across four ecosystems is a reliable way to be mediocre in all of them. The mistake is not concentration. The mistake is concentration you never chose, arrived at one convenient default at a time, and would not have agreed to if someone had shown you the number three years earlier.

What still goes wrong

The concentration number is an estimate, and the weakest input is the one you most want to be confident about. You cannot know how long a rebuild takes until you have done one, and everybody guesses low. Treat the figure as a way to compare this quarter with last quarter rather than as a forecast, and revise it upward the first time a migration you thought was a week turns into a month.

Governance is not permanent either. An open licence and a foundation reduce the odds of a single company changing the rules unilaterally, but they do not eliminate them. The Model Context Protocol’s own governance concentrates final authority in lead maintainers who can veto decisions by the core maintainer group [4], which is a normal and workable arrangement for an open project and still a small number of people. Re-run the four questions in this guide when the people or the licence change, not once at the start.

Finally, this guide optimises for seeing the exposure clearly, which is not the same as removing it, and removing it is often the wrong goal. Some of the most durable small businesses are deeply committed to one ecosystem and are paid for exactly that. The failure this guide is written against is narrower: drifting into a dependency by default, discovering its size only when a price changes or a client asks a question, and finding that the answer was in a document you could have read at any point in the previous three years.